feat(memhub): configurable SDK retrieval and extraction features

MemHub hard-codes none of the SDK's optional features any more. A person
can turn them on from the CLI or the API; with nothing set the SDK config
is byte for byte what it was, so the App behaves as before.

- Features (one global `memory_features` document in config.json, beside
  `memory_mode`; strict booleans, unknown keys refused, absent = SDK
  default):
  - graph_rerank: graph_store.graph_rerank_enabled, usage A, ultra only.
  - same_series_timeline: recall.same_series_timeline_enabled, needs
    graph_rerank.
  - episode_recall: recall.episode_recall_enabled, usage B, pro or ultra.
  - session_episodes: recall.session_episode_enabled, the episode write
    side.
  - episode_summary: recall.episode_summary_enabled, one LLM call per
    closed episode.
  - intention: llm.enable_intention, on by default.
  - few_shot: llm.few_shot_enabled.
- How they reach the SDK: build_sdk_config sets the MemoryConfig fields
  per client. Every exposed flag is a config field read at call time, never
  os.environ; the SDK's own env var still sets the default of an unset
  feature. A feature never touches the vector store, embedder or store
  paths, so collection identity is unchanged.
- A change rebuilds the handles like a mode change: write handles are keyed
  by (profile, revision, mode, features) and retired with the #218 use
  count; the read handle (ActiveProfileReadHandle) is keyed by (profile,
  mode, features) and builds a new client on the next read, closing the old
  one once no request uses it.
- API: GET/PUT /api/v1/settings/memory-features. Each feature reports its
  effective value, configured value, default, whether it is active in the
  current mode (e.g. graph_rerank in pro: requires_ultra), cost, SDK field,
  env var and a description. PUT: a value sets, null resets, an absent key
  is kept, anything else is 422.
- Status: pipeline.features (the non-default ones) in
  /api/v1/memory-processing, memhub status --json and progress --json.
- CLI: memhub settings features [set <name>=<bool> ... [--yes] |
  reset [<name> ...]]; set previews without --yes.
- Contract: docs/contracts/memory-features-v1.md. It has the inventory,
  what is left out and why, and zh/en/es copy.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

.env.example

@@ -40,6 +40,7 @@ MEMORY_EMBEDDING_DIMS=1024
 # MEMORY_EMBEDDER_TIMEOUT=15
 # MEMORY_EMBEDDER_MAX_RETRIES=2
 # MEMORY_EMBEDDER_RETRY_DELAY=0.3
+# MEMORY_EMBEDDER_BATCH_SIZE=64
 # MEMORY_EMBEDDER_EXTRA_HEADERS={}
 # MEMORY_EMBEDDER_EXTRA_BODY={}
 

CHANGELOG.md

@@ -4,6 +4,99 @@ Hy Memory and MemHub are versioned and released independently.
 
 ## Unreleased
 
+## Hy Memory 2.0.10 — 2026-11-01
+
+- New `hy_memory.reembed` module for switching the embedding
+  model. Writers must be paused for the whole run. Nothing is derived from
+  the provider or endpoint: source and target are named explicitly, and the
+  identity of a space is the model name plus the width (`space_identity`).
+  Only Chroma and Kuzu are supported.
+  - `vector_space_names(config, base_name=None)` returns every physical
+    collection a client opens for the six vector roles: memory, entities,
+    coding keys and working facts, each in the working and coding variants.
+  - `reembed_vector_space(source, target, embedding=...)` pages every raw
+    source collection, covering all users and all layers including L1_RAW. It
+    embeds the text the write path embedded (`reembed_text`: the canonical
+    state's `state_retrieval_text`, a schema's `retrieval_text`, otherwise the
+    document). It then writes each record into the target with the same id,
+    document and metadata. Each batch's widths are checked before anything is
+    written. A cancel stops at a batch boundary (`ReembedCancelled`). A rerun
+    skips the ids the target already holds, so it resumes, and a merge never
+    overwrites the target. The source is never written.
+  - `plan_reembed`, `verify_reembed` and `drop_vector_space` complete the
+    vector side. `verify_reembed` checks ids, metadata, per-user counts and
+    the width, and runs a self-recall sample. `drop_vector_space` refuses
+    collections that a live client in this process holds open.
+  - `scan_vector_spaces(persist_directory)` is read-only. It lists every
+    space on disk with its roles, counts, stored width, per-user counts and
+    the models recorded in metadata. `diff_spaces(source, target)` returns the
+    ids that the source holds and the target lacks, whatever their status.
+  - `reembed_graph(source_path, target_path, embedding=..., target_dims=...)`
+    builds a new Kuzu graph at the new width. It copies every node and edge
+    and recomputes `Memory.embedding`, `Memory.beh_embedding`, and the Topic
+    and HierarchyTopic embeddings, then rebuilds the HNSW index. The source
+    graph is opened read-only and kept for rollback. A run that fails or is
+    cancelled leaves no target behind, and a rerun starts over.
+    `verify_graph_reembed` and `drop_graph` complete the graph side.
+- `EmbedService.embed_batch` now splits large batches into
+  requests of `embedder.batch_size` texts each (`MEMORY_EMBEDDER_BATCH_SIZE`,
+  default 64); before, it sent them all in one request. On a 429 it waits for
+  the `Retry-After` time plus jitter, capped at 60 s. It takes a per-call
+  `max_retries`. A returned vector whose width differs from the configured
+  `embedding_dims` now raises `EmbeddingDimensionMismatchError` and is not
+  retried.
+- Raw conversations are stored already trimmed. MemHub now keeps every raw
+  forever and lite mode stores nothing else, so tool output and long pasted
+  logs no longer stay in full. One rule (`hy_memory/utils/raw_trim.py`) is
+  applied to each message when the raw is written:
+  - `user` and `system` turns are kept in full;
+  - `assistant` turns and other speakers (`[John]:`) longer than 500
+    characters, after stripping, are cut to the first 500 plus `...`. This is
+    the limit history injection used before;
+  - `tool` turns keep their head and tail, and the middle becomes
+    `[...N bytes omitted by truncate...]`, with the coding limit
+    `coding.tool_result_max_bytes` (`MEMORY_CODING_TOOL_RESULT_MAX_BYTES`,
+    default 2048).
+
+  The rule is idempotent. It covers chat `L1_RAW` in lite, pro and ultra (lite
+  keeps tool turns, because it has no coding route) and the `memory_history`
+  ADD record. The coding/working `coding_trace` was already trimmed when it
+  was written. Its tool and assistant truncations are now idempotent, so
+  reading the history back no longer rewrites the omitted-count markers. The
+  text embedded for `L1_RAW` is the trimmed text. New raws carry
+  `custom.raw_trimmed=true`.
+- Extractor history ("last k messages") now joins the stored raws directly
+  and does not trim them again. Raws written before this change have no
+  `raw_trimmed` flag, so they are still trimmed with the same rule when they
+  are read. The extractor still receives the current batch in full (tool
+  turns are still removed on the chat route).
+
+## MemHub 1.0.5 — 2026-10-28
+
+Build 36, on `hy-memory==2.0.8`. Linux x64/arm64 Runtime and macOS arm64
+App and Runtime. Upgrading from 1.0.4 migrates the existing library to SDK
+authority on first launch.
+
+- MemHub turns off the HY pipeline log and trace at every diagnostic level
+  (hy-memory 2.0.8 `MEMORY_PIPELINE_LOG_ENABLED=false`,
+  `MEMORY_PIPELINE_TRACE_ENABLED=false`), so the per-step JSONL files under
+  `hy-memory/governed-v2/logs/` and the `pipeline_logs` rows stop growing.
+  The Debug panel no longer shows the HY pipeline section. Existing files are
+  left in place.
+- HY SDK text logs are written once, to `hy_memory.log`. They are no longer
+  copied to `service.log` and `service-error.log` through the SDK's
+  stdout/stderr streams. SDK warnings and errors also go to `core.log`, so the
+  Debug panel still shows them. A live switch of the diagnostic level now also
+  changes the level of `hy_memory.log`.
+
+- MemHub shows when working and coding experiences were talked about. After
+  the SDK upgrade every one of them read "时间未知" (time unknown) because the
+  SDK never stored their conversation time; hy-memory 2.0.7 stores it and
+  Core now passes it through the memory list, detail, search and home
+  summary. An experience stored before 2.0.7 shows its write time, labelled
+  "写入于" / "写入时间", instead of "时间未知". Metadata repair also writes
+  the conversation time it recovers from the migration bundle into the SDK's
+  coding/working stores when they already have the column.
 - The SDK migration no longer fails its first run on a home with legacy
   vectors: reattaching them left a Chroma instance open with its own settings,
   and the readback that followed could not open the same store
@@ -68,6 +161,67 @@ Hy Memory and MemHub are versioned and released independently.
   `hy_node_permissions` proof table. Evidence, governed memories, model
   profiles, SDK stores, and source/host grants are preserved.
 
+## Hy Memory 2.0.9 — 2026-11-01
+
+- Published
+  [`hy-memory==2.0.9`](https://pypi.org/project/hy-memory/2.0.9/).
+- Placeholder values no longer overwrite the user's basic profile (L0). A
+  small extraction model filled `basic_info` with `{"name": "用户"}`,
+  `The user` or `N/A` when the user had said nothing about themselves, and
+  each placeholder superseded the real value, so the L0 chain read
+  费天翔 → Fei → 用户 → The user. Two defences:
+  - `upsert_basic_profile` drops placeholder values: empty, `用户`,
+    `the user`, `user`, `n/a`, `na`, `none`, `null`, `unknown`, `未提供`,
+    `未知`, `不详`, `-`, `?` and the field's own name, compared
+    case-insensitively with whitespace and punctuation stripped. When nothing
+    is left it writes no new version and no new L0 head. Placeholders already
+    stored in the chain no longer override earlier real values when the
+    profile is assembled, so `get_user_profile` of an affected user shows the
+    last real value again without any migration. This also stops an
+    assistant quoting an old memory ("The user's name is 用户.") from being
+    written back.
+  - The extraction prompts (zh and en) state the rule: only what the user
+    said in a `user` turn, never the field name or a placeholder, `{}`
+    otherwise, with a small-talk counter-example.
+
+## Hy Memory 2.0.8 — 2026-10-28
+
+- Published
+  [`hy-memory==2.0.8`](https://pypi.org/project/hy-memory/2.0.8/).
+- New `MEMORY_PIPELINE_LOG_ENABLED` (default `true`) turns the pipeline JSONL
+  log (`{MEMORY_LOG_DIR}/<subdir>/<date>.log`) off. With `false` no JSONL
+  file or directory is written. When `MEMORY_PIPELINE_TRACE_ENABLED` is also
+  `false`, `store_pipeline_log` becomes a no-op that returns `True` at once:
+  no timing, redaction, background task or database write. The init log line
+  now states whether the JSONL log is on or off.
+- Conversation and memory text no longer appears in INFO or WARNING logs. The
+  `[add]` input, search queries and rewrites, coding experience tasks, fact
+  statements and subjects, reconcile op details, extracted info, embedding
+  inputs and unparsable model output moved to DEBUG. INFO lines keep ids,
+  counts and character lengths; parse-failure warnings remain but report only
+  the length of the output, e.g. `failed to parse output (812 chars)`. The
+  `[embed]` INFO line now carries `request.input_chars` instead of
+  `request.input`. The `JSON_PARSE_FAILED` error message from extraction no
+  longer embeds the model output either.
+
+## Hy Memory 2.0.7 — 2026-10-28
+
+- Published
+  [`hy-memory==2.0.7`](https://pypi.org/project/hy-memory/2.0.7/).
+- Coding and working experiences and facts now keep their conversation time.
+  `add(..., memory_at=...)` reached only the chat path before; the
+  productivity path dropped it, so every experience had only its write time
+  and MemHub showed them as "时间未知" (time unknown). `coding_memory_meta`
+  and `working_fact` gain a nullable `memory_at` column (epoch seconds) in
+  both the working and coding stores; existing databases get it on open, and
+  their old rows stay `None`. Without `memory_at` a write uses the time the
+  request arrived, as chat does. When an experience is updated by a later
+  conversation it keeps the later of the two times.
+- `list_memories` (`scene="coding"`, `"working"` and `"all"`), `get` and
+  `search` return `memory_at` for experiences and facts in the same field and
+  format as chat (epoch seconds, `None` when unknown). Experience search
+  results now also carry `created_at`.
+
 ## Hy Memory 2.0.6 — 2026-10-27
 
 - Published

README.md

@@ -33,6 +33,7 @@ English | [中文](README.zh.md)
 ## Latest News
 
 * *<mark>[new]</mark>* 🔥🔥🔥 **[2026-10]**: 🚀 **Open-sourced** quantized **FP8** (**[Hy-Memory-30B-A3B-FP8](https://huggingface.co/tencent/Hy-Memory-30B-A3B-FP8)**) and **GGUF** (**[Hy-Memory-30B-A3B-GGUF](https://huggingface.co/tencent/Hy-Memory-30B-A3B-GGUF)**) versions of **Hy-Memory-30B-A3B** on Hugging Face!
+* *<mark>[new]</mark>* 🔥🔥🔥 **[2026-10]**: 📦 **Released [MemHub v1.0.5](https://github.com/Tencent-Hunyuan/Hy-Memory/releases/tag/v1.0.5)** with Chinese, English and Spanish interfaces, one-time pairing codes for more custom Harnesses, and smoother browsing, search, startup and recovery, on **`hy-memory==2.0.8`**. See the release notes for details.
 * *<mark>[new]</mark>* 🔥🔥🔥 **[2026-10]**: 📦 **Released [MemHub v1.0.4](https://github.com/Tencent-Hunyuan/Hy-Memory/releases/tag/v1.0.4)** and **[memhub-cli 1.0.5](https://www.npmjs.com/package/memhub-cli/v/1.0.5)**, with a signed macOS desktop app and verified macOS/Linux Runtimes!
 * *<mark>[new]</mark>* 🔥🔥🔥 **[2026-10]**: 🚀 **Released** **Hy Memory 2.0 (`hy-memory==2.0.4`)** on [PyPI](https://pypi.org/project/hy-memory/2.0.4/) and open-sourced the **Hy-Memory** monorepo together with **[MemHub](memhub/README.md)**!
 * *<mark>[new]</mark>* 🔥🔥🔥 **[2026-10]**: 🚀 **Open-sourced** **[Hy-Memory-30B-A3B](https://huggingface.co/tencent/Hy-Memory-30B-A3B)** on Hugging Face along with its [technical report](https://arxiv.org/abs/2610.22196) on attributed reinforcement learning for long-term memory writing!

README.zh.md

@@ -33,6 +33,7 @@
 ## 最新动态
 
 * *<mark>[new]</mark>* 🔥🔥🔥 **[2026-10]**: 🚀 **模型开源** 在 Hugging Face 发布 **Hy-Memory-30B-A3B** 的 **FP8**（**[Hy-Memory-30B-A3B-FP8](https://huggingface.co/tencent/Hy-Memory-30B-A3B-FP8)**）与 **GGUF**（**[Hy-Memory-30B-A3B-GGUF](https://huggingface.co/tencent/Hy-Memory-30B-A3B-GGUF)**）量化版本模型！
+* *<mark>[new]</mark>* 🔥🔥🔥 **[2026-10]**: 📦 **正式发布 [MemHub v1.0.5](https://github.com/Tencent-Hunyuan/Hy-Memory/releases/tag/v1.0.5)**：支持中文、英文和西班牙语界面，可通过一次性配对码连接更多自定义 Harness，并优化了记忆浏览、搜索、启动和异常恢复；底层使用 **`hy-memory==2.0.8`**。详见发布说明。
 * *<mark>[new]</mark>* 🔥🔥🔥 **[2026-10]**: 📦 **正式发布 [MemHub v1.0.4](https://github.com/Tencent-Hunyuan/Hy-Memory/releases/tag/v1.0.4)** 与 **[memhub-cli 1.0.5](https://www.npmjs.com/package/memhub-cli/v/1.0.5)**，提供已签名的 macOS 桌面应用及经过验证的 macOS/Linux Runtime！
 * *<mark>[new]</mark>* 🔥🔥🔥 **[2026-10]**: 🚀 **正式发布** **Hy Memory 2.0 (`hy-memory==2.0.4`)**（已上线 [PyPI](https://pypi.org/project/hy-memory/2.0.4/)），并开源 **Hy-Memory** monorepo 与 **[MemHub](memhub/README.md)**！
 * *<mark>[new]</mark>* 🔥🔥🔥 **[2026-10]**: 🚀 **模型开源** 在 Hugging Face 发布 **[Hy-Memory-30B-A3B](https://huggingface.co/tencent/Hy-Memory-30B-A3B)** 长期记忆写入模型及[技术报告](https://arxiv.org/abs/2610.22196)！

docs/env_reference.md

@@ -48,6 +48,7 @@
 | `MEMORY_EMBEDDER_TIMEOUT` | `15` | 单次超时，秒 |
 | `MEMORY_EMBEDDER_MAX_RETRIES` | `2` | 最大重试次数 |
 | `MEMORY_EMBEDDER_RETRY_DELAY` | `0.3` | 重试间隔，秒 |
+| `MEMORY_EMBEDDER_BATCH_SIZE` | `64` | `embed_batch` 单次请求最多发送的文本条数，超出按此切块 |
 | `MEMORY_EMBEDDER_EXTRA_HEADERS` | — | JSON object，附加 HTTP headers |
 | `MEMORY_EMBEDDER_EXTRA_BODY` | — | JSON object，附加请求 body |
 
@@ -191,6 +192,7 @@ SDK 不设置 Rerank Search 并发上限。开启后模型在 `HyMemoryClient`
 | `MEMORY_LOG_FORMAT` | `text` | `text` / `json` |
 | `MEMORY_LOG_DIR` | — | 统一日志根目录 |
 | `MEMORY_TRACE_LOG_DIR` | — | JSONL 日志目录的兼容旧名；优先使用 `MEMORY_LOG_DIR` |
+| `MEMORY_PIPELINE_LOG_ENABLED` | `true` | 是否写 pipeline JSONL 日志（`{MEMORY_LOG_DIR}/<subdir>/<date>.log`）；`false` 完全不写 |
 | `MEMORY_PIPELINE_TRACE_ENABLED` | `true` | 是否写入 cache 的 pipeline trace 表 |
 | `MEMORY_MEMORY_OPERATIONS_ENABLED` | `false` | 是否写入 memory operation 审计记录 |
 | `MEMORY_HISTORY_ENABLE` | `true` | 是否启用历史记录 |

hy_memory/__init__.py

@@ -55,6 +55,33 @@ except Exception:
 from .client import HyMemoryClient
 from .inspector import MemoryInspector
 from .runtime import SharedRuntime
+from .reembed import (
+    EmbeddingSpaceIdentity,
+    GraphReembedPlan,
+    GraphReembedProgress,
+    GraphReembedReport,
+    GraphVerification,
+    ReembedCancelled,
+    ReembedPlan,
+    ReembedProgress,
+    ReembedReport,
+    ReembedVerification,
+    ScannedSpace,
+    SpaceDiff,
+    VectorSpace,
+    diff_spaces,
+    drop_graph,
+    drop_vector_space,
+    plan_reembed,
+    reembed_graph,
+    reembed_text,
+    reembed_vector_space,
+    scan_vector_spaces,
+    space_identity,
+    vector_space_names,
+    verify_graph_reembed,
+    verify_reembed,
+)
 from .snapshot import (
     DataPlaneRestore,
     DataPlaneSnapshot,
@@ -129,6 +156,32 @@ __all__ = [
     "DataPlaneSnapshotRejected",
     "restore_local_data_plane",
     "snapshot_local_data_plane",
+    # 换 embedding 模型：把向量库和图重新 embed 进新的 embedding 空间
+    "EmbeddingSpaceIdentity",
+    "GraphReembedPlan",
+    "GraphReembedProgress",
+    "GraphReembedReport",
+    "GraphVerification",
+    "ReembedCancelled",
+    "ReembedPlan",
+    "ReembedProgress",
+    "ReembedReport",
+    "ReembedVerification",
+    "ScannedSpace",
+    "SpaceDiff",
+    "VectorSpace",
+    "diff_spaces",
+    "drop_graph",
+    "drop_vector_space",
+    "plan_reembed",
+    "reembed_graph",
+    "reembed_text",
+    "reembed_vector_space",
+    "scan_vector_spaces",
+    "space_identity",
+    "vector_space_names",
+    "verify_graph_reembed",
+    "verify_reembed",
     # 配置
     "MemoryConfig",
     # 公共扩展协议 / 精确读取类型

hy_memory/agent/extractor.py

@@ -682,6 +682,23 @@ Output:
 ```
 John's occupation and location are HIS attributes, so they go into `memory` with "John" as the subject. They must NEVER land in `basic_info.occupation` / `basic_info.location` — `basic_info` is exclusively the user's own profile. Only `name` (Tim, from the user message's `name` field) belongs there.
 
+## Example 11 — Counter-example: small talk has no basic_info
+
+New Messages:
+[{"role": "user", "content": "Hi"},
+ {"role": "assistant", "content": "Hello! The user's name is The user. How can I help?"}]
+Memory Date: 2025-09-03
+
+Output:
+```json
+{
+  "memory": [],
+  "intentions": [],
+  "basic_info": {}
+}
+```
+The user stated nothing about themselves. The assistant's line is not a user turn, and "The user" is a placeholder, not a name. Output `{}` — never `{"name": "The user"}` or `{"name": "user"}`.
+
 ---
 """
 
@@ -749,6 +766,14 @@ Rules:
   fields (e.g. name, age, location, occupation, employer) **explicitly stated**
   in the NEW user messages. Never infer, never invent. If no field is explicitly
   stated, output `{{}}`.
+  * HARD RULE: fill a field ONLY if the user stated it in a `user` turn.
+    Anything said only by the assistant (including the assistant quoting an
+    earlier memory such as "The user's name is X") does NOT count. If nothing
+    qualifies, output `"basic_info": {{}}`.
+  * NEVER write the field name itself or a placeholder as a value ("user",
+    "The user", "用户", "name", "N/A", "unknown", "none", "-"). Omit the field.
+  * Negative example: small talk such as `[user]: Hi` / `[assistant]: How can
+    I help?`, or `[user]: Reply with exactly: PONG` → `"basic_info": {{}}`.
   * Exception for `name`: the `name` field on the user's own message, or a leading
     line "The user's name is X" / "用户名叫X", is an authoritative source for
     `name` — even though the user never said it out loud.
@@ -817,6 +842,14 @@ Rules:
   fields (e.g. name, age, location, occupation, employer) **explicitly stated**
   in the NEW user messages. Never infer, never invent. If no field is explicitly
   stated, output `{{}}`.
+  * HARD RULE: fill a field ONLY if the user stated it in a `user` turn.
+    Anything said only by the assistant (including the assistant quoting an
+    earlier memory such as "The user's name is X") does NOT count. If nothing
+    qualifies, output `"basic_info": {{}}`.
+  * NEVER write the field name itself or a placeholder as a value ("user",
+    "The user", "用户", "name", "N/A", "unknown", "none", "-"). Omit the field.
+  * Negative example: small talk such as `[user]: Hi` / `[assistant]: How can
+    I help?`, or `[user]: Reply with exactly: PONG` → `"basic_info": {{}}`.
   * Exception for `name`: the `name` field on the user's own message, or a leading
     line "The user's name is X" / "用户名叫X", is an authoritative source for
     `name` — even though the user never said it out loud.
@@ -1177,7 +1210,7 @@ class Extractor:
                         if not tc_id:
                             logger.error(
                                 f"[extractor] tool_call_id is empty! "
-                                f"raw tc: {json.dumps(tc_raw, ensure_ascii=False, default=str)[:500]}"
+                                f"raw_keys={list(tc_raw.keys())}"
                             )
                         try:
                             t_res = await tool_registry.dispatch(
@@ -1311,11 +1344,12 @@ class Extractor:
                 if parsed is None:
                     if not all_tool_calls:
                         logger.warning(
-                            f"[extractor] JSON parse failed: {raw_text[:200]}"
+                            f"[extractor] JSON parse failed ({len(raw_text)} chars)"
                         )
+                        logger.debug(f"[extractor] unparsable output: {raw_text[:200]!r}")
                         return ExtractResult(
                             success=False,
-                            error=f"JSON parse failed: {raw_text[:200]}",
+                            error=f"JSON parse failed ({len(raw_text)} chars)",
                             error_code="JSON_PARSE_FAILED",
                             raw_response=raw_text,
                             tokens_used=total_tokens,

hy_memory/agent/llm_provider.py

@@ -129,8 +129,7 @@ def _parse_tool_calls_list(raw_calls: list, _json) -> List[Dict[str, Any]]:
         tc_id = tc.get("id") or ""
         if not tc_id:
             logger.error(
-                f"[llm] tool_call missing 'id'! raw: "
-                f"{_json.dumps(tc, ensure_ascii=False, default=str)[:500]}"
+                f"[llm] tool_call missing 'id'! keys={list(tc.keys())}"
             )
 
         # 保留原始 tc 完整结构，只规范化 arguments

hy_memory/agent/mem_agent.py

@@ -234,7 +234,7 @@ class MemAgent:
                     f"identity={len(identity)} facts={len(facts)} "
                     f"basic_info={basic_field_count}fields"
                 )
-                logger.info(f"MemAgent: extracted_info={extract_result.info}")
+                logger.debug(f"MemAgent: extracted_info={extract_result.info}")
 
             # 3. 冲突检测（如果有已有记忆）
             if mode == ProcessMode.FULL and existing_memories:

hy_memory/agent/prompts_zh.py

@@ -365,6 +365,13 @@ EXTRACT_PROMPT_ZH = """## 最近 k 条消息
 * basic_info：**只装用户本人的画像**。
   仅当 NEW user messages 中**明确**陈述稳定画像字段（如 姓名、年龄、所在地、职业、雇主）时才填写。
   绝不推断、绝不编造。没有任何字段被明确陈述时输出 `{{}}`。
+  * **硬性规则**：只有用户在 `user` 回合里**明确说出**的字段才能填写。只在
+    assistant 回合出现的内容（包括 assistant 引用旧记忆，如"The user's name is X"
+    / "用户名是X"）一律不算。没有符合条件的字段时，必须输出 `"basic_info": {{}}`。
+  * **绝不**把字段名本身或占位符当值写入（"用户"、"The user"、"user"、"name"、
+    "姓名"、"N/A"、"unknown"、"未知"、"未提供"、"-"），宁可省略该字段。
+  * 负例：寒暄如 `[user]: 你好` / `[assistant]: 我可以帮你。`，或
+    `[user]: Reply with exactly: PONG` → `"basic_info": {{}}`。
   * `name` 的例外：用户自己那条消息上的 `name` 字段，或文首形如
     "用户名叫X" / "The user's name is X" 的那一行，都是 `name` 的权威来源
     ——即使用户本人从未开口自报姓名。
@@ -440,6 +447,13 @@ EXTRACT_PROMPT_ZH_NO_INTENTION = """## 最近 k 条消息
 * basic_info：**只装用户本人的画像**。
   仅当 NEW user messages 中**明确**陈述稳定画像字段（如 姓名、年龄、所在地、职业、雇主）时才填写。
   绝不推断、绝不编造。没有任何字段被明确陈述时输出 `{{}}`。
+  * **硬性规则**：只有用户在 `user` 回合里**明确说出**的字段才能填写。只在
+    assistant 回合出现的内容（包括 assistant 引用旧记忆，如"The user's name is X"
+    / "用户名是X"）一律不算。没有符合条件的字段时，必须输出 `"basic_info": {{}}`。
+  * **绝不**把字段名本身或占位符当值写入（"用户"、"The user"、"user"、"name"、
+    "姓名"、"N/A"、"unknown"、"未知"、"未提供"、"-"），宁可省略该字段。
+  * 负例：寒暄如 `[user]: 你好` / `[assistant]: 我可以帮你。`，或
+    `[user]: Reply with exactly: PONG` → `"basic_info": {{}}`。
   * `name` 的例外：用户自己那条消息上的 `name` 字段，或文首形如
     "用户名叫X" / "The user's name is X" 的那一行，都是 `name` 的权威来源
     ——即使用户本人从未开口自报姓名。
@@ -669,6 +683,23 @@ Memory Date：2025-09-02
 ```
 职业和所在地是 **John 的**属性，所以进 `memory` 并以"John"作主体，**绝不能**落到 `basic_info` 的 `occupation` / `location` —— `basic_info` 专属于用户本人。只有 `name`（Tim，来自用户消息的 `name` 字段）才该放进去。注意输出语言随输入语言（此例为英文）。
 
+## 示例 11 —— 反例：寒暄没有 basic_info
+
+新消息：
+[{"role": "user", "content": "你好"},
+ {"role": "assistant", "content": "你好！我可以帮你。之前记得 The user's name is 用户."}]
+Memory Date：2025-09-03
+
+输出：
+```json
+{
+  "memory": [],
+  "intentions": [],
+  "basic_info": {}
+}
+```
+用户没有说任何关于自己的信息。assistant 那句不是 user 回合，"用户" 也是占位符而不是名字。必须输出 `{}`——**绝不能**输出 `{"name": "用户"}`。
+
 ---
 """
 

hy_memory/agent/reconciler.py

@@ -50,7 +50,7 @@ def is_canonical_state_node(node: Any) -> bool:
 
 
 def _log_ops_detail(ops: List["ReconcileOp"]) -> None:
-    """以单条 INFO（JSON list）输出 reconcile 的逐条结果，便于线上 debug。"""
+    """以单条 DEBUG（JSON list）输出 reconcile 的逐条结果；含记忆正文，不进 INFO。"""
     if not ops:
         return
     view = [
@@ -68,7 +68,7 @@ def _log_ops_detail(ops: List["ReconcileOp"]) -> None:
         }
         for o in ops
     ]
-    logger.info(
+    logger.debug(
         f"[reconciler] ops detail: {json.dumps(view, ensure_ascii=False, default=str)}"
     )
 
@@ -1514,9 +1514,10 @@ class MemoryReconciler:
                 logger.warning(
                     f"[reconciler] RECONCILE JSON parse failed "
                     f"(attempt {_attempt+1}/{_MAX_PARSE_ATTEMPTS}), "
-                    f"retrying: raw="
-                    f"{raw_response[:200] if _log_llm_content_enabled() else '<redacted>'}"
+                    f"retrying: raw({len(raw_response or '')} chars)"
                 )
+                if _log_llm_content_enabled():
+                    logger.debug(f"[reconciler] unparsable raw={raw_response[:200]!r}")
 
             recon_elapsed = (_time.perf_counter() - t_recon) * 1000
 
@@ -1528,8 +1529,8 @@ class MemoryReconciler:
                     error_code = "LLM_UNAVAILABLE"
                 elif raw_response.strip():
                     error_msg = (
-                        f"RECONCILE JSON parse failed after {_n_attempts} attempts: "
-                        f"{raw_response[:200] if _log_llm_content_enabled() else '<redacted>'}"
+                        f"RECONCILE JSON parse failed after {_n_attempts} attempts "
+                        f"({len(raw_response)} chars)"
                     )
                     error_code = "JSON_PARSE_FAILED"
                 else:
@@ -1806,7 +1807,8 @@ class MemoryReconciler:
                 except Exception as _re:
                     logger.debug(f"[reconciler] json_repair failed: {_re}")
             if data is None:
-                logger.warning(f"[reconciler] failed to parse ops JSON: {text[:200]}")
+                logger.warning(f"[reconciler] failed to parse ops JSON ({len(text)} chars)")
+                logger.debug(f"[reconciler] unparsable ops output: {text[:200]!r}")
                 return []
 
         if not isinstance(data, list):
@@ -2018,11 +2020,11 @@ class MemoryReconciler:
                 if clash is not None:
                     logger.warning(
                         f"[reconciler] memory_id {clash} already touched by '{touched[clash]}', "
-                        f"current op '{op.op} ({op.reason[:60]})' dropped"
+                        f"current op '{op.op}' dropped"
                     )
                     continue
                 for t in target_ids:
-                    touched[t] = f"{op.op}({op.reason[:40]})"
+                    touched[t] = op.op
             validated.append(op)
 
         return validated

hy_memory/agent/tools/base.py

@@ -211,7 +211,9 @@ def parse_tool_calls_from_json(raw: Any) -> List[ToolCall]:
                 try:
                     args = json.loads(args_raw) if args_raw.strip() else {}
                 except json.JSONDecodeError:
-                    logger.warning(f"[tools] failed to parse arguments as JSON: {args_raw[:120]}")
+                    logger.warning(
+                        f"[tools] failed to parse arguments as JSON ({len(args_raw)} chars)"
+                    )
                     args = {}
             elif isinstance(args_raw, dict):
                 args = args_raw

hy_memory/agent/tools/basic_profile.py

@@ -26,6 +26,8 @@ from __future__ import annotations
 
 import asyncio
 import logging
+import re
+import unicodedata
 import uuid
 import weakref
 from dataclasses import dataclass
@@ -208,6 +210,76 @@ def _sanitize_kv(
     return result
 
 
+# ================================================================
+# 占位符守卫（P0-1）
+# ================================================================
+# 弱模型（如 qwen3.5-2b）在用户什么都没说时，会把 schema 字段说明里的词当值
+# 填进 basic_info：{"name": "用户"} / "The user" / "N/A"。每个占位符都会
+# supersede 掉真实值，演化链变成 费天翔 → Fei → 用户 → The user。
+# 规则：占位符永远不入库；读侧组装 KV 时占位符也不覆盖已有真实值。
+
+_PLACEHOLDER_VALUES = frozenset({
+    "用户", "该用户", "本用户", "用户本人",
+    "the user", "the user's", "user", "a user", "this user",
+    "n/a", "na", "none", "null", "nil", "unknown", "unspecified",
+    "not provided", "not specified", "not mentioned",
+    "未提供", "未知", "不详", "未说明", "未提及", "无",
+    "-", "?",
+})
+
+# 内置字段的中文标签：弱模型也会把字段说明的中文词当值。
+_FIELD_LABELS: Dict[str, tuple] = {
+    "name": ("姓名", "名字", "用户名", "用户姓名", "用户名字"),
+    "age": ("年龄", "用户年龄"),
+    "location": ("所在地", "地点", "城市", "用户所在地"),
+    "occupation": ("职业", "工作", "用户职业"),
+    "employer": ("雇主", "公司", "用户雇主"),
+}
+
+# 去掉标点/符号后再比较，让 "The user." / "用户。" / "N/A" 都命中。
+# "-" / "?" 本身就是纯标点，剥完为空串，同样判为占位符。
+_PUNCT_RE = re.compile(r"[\s\.,;:!?'\"`~()\[\]{}<>/\\|_\-\u3000-\u303f\uff00-\uff0f\uff1a-\uff20]+")
+
+
+def _placeholder_key(value: str) -> str:
+    s = unicodedata.normalize("NFKC", str(value or "")).strip().lower()
+    s = s.replace("\u2019", "'")
+    s = _PUNCT_RE.sub(" ", s)
+    return " ".join(s.split())
+
+
+_PLACEHOLDER_KEYS = frozenset(_placeholder_key(v) for v in _PLACEHOLDER_VALUES)
+
+
+def is_placeholder_value(field_name: str, value: Any) -> bool:
+    """value 是否为占位符（空 / 用户 / The user / N/A / 字段名本身 等）。
+
+    大小写不敏感、首尾空白与标点不敏感。数字（如 age=30）永远不是占位符。
+    """
+    if value is None:
+        return True
+    if isinstance(value, bool):
+        return False
+    if isinstance(value, (int, float)):
+        return False
+    key = _placeholder_key(value)
+    if not key:
+        return True
+    if key in _PLACEHOLDER_KEYS:
+        return True
+    field_key = _placeholder_key(field_name)
+    if field_key and key in (field_key, field_key.replace("_", " ")):
+        return True
+    for label in _FIELD_LABELS.get(field_name, ()):
+        if key == _placeholder_key(label):
+            return True
+    return False
+
+
+def _drop_placeholders(kv: Dict[str, Any]) -> Dict[str, Any]:
+    return {k: v for k, v in kv.items() if not is_placeholder_value(k, v)}
+
+
 # ================================================================
 # 公共入口：upsert_basic_profile
 # ================================================================
@@ -287,6 +359,8 @@ async def _upsert_basic_profile_unlocked(
     # L0 Basic Info 从输入、唯一性到落盘都只属于 user_id。
 
     new_kv = _sanitize_kv(kv or {}, allowed_fields=allowed_fields)
+    # 单调性守卫：占位符永远不覆盖已有值，过滤后为空就不写新版本、不产生新链头。
+    new_kv = _drop_placeholders(new_kv)
     if not new_kv:
         return BasicProfileUpsertResult(
             success=True, changed=False, reason="no valid fields after sanitize"
@@ -513,5 +587,8 @@ def _assemble_full_kv_from_nodes(
         for k, v in kv.items():
             if allowed_set is not None and k not in allowed_set:
                 continue
+            # 历史上已落盘的占位符版本（如 name=用户）不覆盖更早的真实值。
+            if is_placeholder_value(k, v):
+                continue
             full_kv[k] = v
     return full_kv

hy_memory/client.py

@@ -744,7 +744,7 @@ class HyMemoryClient:
     def _install_pipeline_observability_hooks(self) -> None:
         """
         包装 cache.store_pipeline_log：
-        - Log：JSONL 文件，始终写入（环节级，不依赖 Trace 开关）
+        - Log：JSONL 文件，由 MEMORY_PIPELINE_LOG_ENABLED 控制（默认开启）
         - Trace：SQLite pipeline_logs，由 MEMORY_PIPELINE_TRACE_ENABLED 控制（Inspector）
         - 落盘 fire-and-forget，不阻塞 search/write 热路径
         """
@@ -958,7 +958,7 @@ class HyMemoryClient:
                 parsed_tool_calls = json.loads(raw_tool_calls) if raw_tool_calls else []
             except Exception:
                 logger.warning(
-                    "tool_calls 是非法 JSON 字符串，已忽略: %r", raw_tool_calls[:200]
+                    "tool_calls 是非法 JSON 字符串，已忽略 (%d chars)", len(raw_tool_calls)
                 )
                 parsed_tool_calls = []
             raw_tool_calls = parsed_tool_calls if isinstance(parsed_tool_calls, list) else []
@@ -1544,6 +1544,7 @@ class HyMemoryClient:
                 # LLM 写的"这个结论一路怎么改成现在这样"，直接进提示词；
                 # 没被覆盖过就是空串。不参与检索——召回认的是 statement 的当前值。
                 "fact_chain_summary": fact.fact_chain_summary,
+                "memory_at": _epoch_seconds(fact.memory_at),
                 "created_at": fact.created_at.isoformat() if fact.created_at else None,
                 "score": round(float(score), 4),
                 "source": fact.source,
@@ -1652,6 +1653,8 @@ class HyMemoryClient:
                 "score": (round(float(score_by_id[d.experience_id]), 4)
                           if d.experience_id in score_by_id else None),
                 "confidence": d.confidence,
+                "memory_at": _epoch_seconds(d.memory_at),
+                "created_at": d.created_at.isoformat() if d.created_at else None,
                 "updated_at": d.updated_at.isoformat() if d.updated_at else None,
                 "source": d.source,
                 "extraction_source": d.extraction_source,
@@ -1953,6 +1956,7 @@ class HyMemoryClient:
                   "content"    chat 正文；经验为 "task\nsolution"；事实为 statement
                   "score"      RRF 融合分；"lane_score" 为该路原始分；"lane_rank" 为该路名次
                   "source", "created_at", "updated_at"（秒级 epoch，取不到为 None）
+                  "memory_at"  对话时间（秒级 epoch），各路原有字段，三路同名同格式
                 chat 路内部名次：normal → schema → proactive 依次排；working/coding 路
                 内部名次：事实与经验各按原顺序交替（事实在前），两层谁也不压谁。
         """
@@ -2144,7 +2148,8 @@ class HyMemoryClient:
                           "limit", "offset", "request_id", "elapsed_ms"}
                 每条带 "scene"、"kind"（chat | experience）、"memory_id"、"content"
                 （经验为 "task\nsolution"）、"source"、"status"、"created_at"、
-                "updated_at"（秒级 epoch），以及该路原有字段。
+                "updated_at"（秒级 epoch），以及该路原有字段（含对话时间
+                "memory_at"，秒级 epoch，三路同名同格式）。
         """
         return self._loop_thread.run(
             self.async_list_memories(
@@ -2967,6 +2972,7 @@ class HyMemoryClient:
                     session_id=session_id,
                     request_id=request_id,
                     source=source,
+                    memory_at=memory_at,
                 )
                 # Public receipt always identifies the productivity namespace.
                 resp["variant"] = scene_variant
@@ -3002,13 +3008,13 @@ class HyMemoryClient:
                 f"[add] writing {preview} user={uid} "
                 f"agent_id={agent_id} session_id={session_id} mode={self._mode}"
             )
-            logger.info(f"[add] messages detail: {[m.to_dict() for m in chat_messages]}")
+            logger.debug(f"[add] messages detail: {[m.to_dict() for m in chat_messages]}")
         else:
             logger.info(
                 f"[add] writing text({len(content)} chars) user={uid} "
-                f"agent_id={agent_id} session_id={session_id} mode={self._mode} "
-                f"content={content}"
+                f"agent_id={agent_id} session_id={session_id} mode={self._mode}"
             )
+            logger.debug(f"[add] text content={content!r}")
 
         # mode → agent_mode 映射
         # lite: 不调 LLM agent
@@ -3083,9 +3089,18 @@ class HyMemoryClient:
             if self._history_store is not None:
                 try:
                     isolation_key = self._build_isolation_key(uid, agent_id)
-                    # 从 content 或 messages 中提取用于记录的文本
+                    # 从 content 或 messages 中提取用于记录的文本。messages 与 L1_RAW
+                    # 同一规则先裁剪再落审计库（工具输出 / 长回复不原样长期保存）。
+                    from .utils.raw_trim import trim_raw_messages
                     history_content = content or "; ".join(
-                        f"{m.role}: {m.content}" for m in chat_messages
+                        f"{m.role}: {m.content}"
+                        for m in trim_raw_messages(
+                            chat_messages,
+                            tool_max_bytes=getattr(
+                                getattr(self._config, "coding", None),
+                                "tool_result_max_bytes", None,
+                            ),
+                        )
                     )
                     await self._history_store.record_add(
                         memory_id=result.memory_id,
@@ -4028,14 +4043,14 @@ class HyMemoryClient:
             decided_variant = HyMemoryClient._scene_variant(scene)
             logger.info(
                 f"[search] scene explicit={decided_scene} (no LLM judge); "
-                f"target={effective_target_query!r}"
+                f"target({len(effective_target_query or '')} chars)"
             )
         elif not agentic_enabled and not queries_supplied:
             # 默认 light search 仅指 query="..."：固定 normal，绝不调用 LLM。
             decided_scene = "normal"
             logger.info(
                 f"[search] search_mode=default single-query scene=normal (no LLM); "
-                f"target={target_query!r}"
+                f"target({len(target_query or '')} chars)"
             )
         elif agentic_enabled:
             # agentic：单/多 query 都强制一次 LLM 完成三分类 + rewrite + 时间解析。
@@ -4070,14 +4085,18 @@ class HyMemoryClient:
                 rewrite_created_after = ar.get("created_after")
                 logger.info(
                     f"[search] agentic scene={decided_scene} via rewrite-LLM "
-                    f"(classify={want_classify}); orig={target_query!r} "
-                    f"effective={effective_target_query!r} rewrite_used={rewrite_used} "
+                    f"(classify={want_classify}); rewrite_used={rewrite_used} "
                     f"created_after={rewrite_created_after}"
                 )
+                logger.debug(
+                    f"[search] agentic rewrite: orig={target_query!r} "
+                    f"effective={effective_target_query!r}"
+                )
             else:
                 decided_scene = "normal"
                 logger.info(
-                    f"[search] agentic rewrite-LLM fallback=normal; target={target_query!r}"
+                    f"[search] agentic rewrite-LLM fallback=normal; "
+                    f"target({len(target_query or '')} chars)"
                 )
         elif rewrite_enabled and not _too_long:
             # default + queries= 的原逻辑：配置开启时一次 LLM 做 rewrite/time，
@@ -4156,7 +4175,7 @@ class HyMemoryClient:
             decided_scene = "normal"
             logger.info(
                 f"[search] default multi-query features disabled; no LLM; "
-                f"target={target_query!r}"
+                f"target({len(target_query or '')} chars)"
             )
 
         # ────────────────────────────────────────────────────────────
@@ -4209,7 +4228,7 @@ class HyMemoryClient:
             else:
                 logger.info(
                     f"[search] routing to PRODUCTIVITY path variant={decided_variant}: "
-                    f"target={effective_target_query!r} "
+                    f"target({len(effective_target_query or '')} chars) "
                     f"workspace_id={workspace_id!r} branch={branch!r}"
                 )
                 return await self._search_coding(
@@ -4237,9 +4256,11 @@ class HyMemoryClient:
         # Chat 召回链路（原有逻辑）
         # ────────────────────────────────────────────────────────────
         logger.info(
-            f"[search] query='{effective_target_query}' user_ids={effective_user_ids} "
+            f"[search] query({len(effective_target_query or '')} chars) "
+            f"user_ids={effective_user_ids} "
             f"agent_ids={effective_agent_ids} session_ids={effective_session_ids} limit={limit}"
         )
+        logger.debug(f"[search] query={effective_target_query!r}")
 
         # rewrite 抽出的时间下界与用户显式 created_after 取更晚者（更严格）
         effective_created_after = created_after
@@ -5920,6 +5941,7 @@ class HyMemoryClient:
                 "source_fact_ids": list(d.source_fact_ids),
                 "confidence": d.confidence,
                 "status": MemoryStatus.ACTIVE.value,
+                "memory_at": _epoch_seconds(d.memory_at),
                 "created_at": d.created_at.isoformat() if d.created_at else None,
                 "updated_at": d.updated_at.isoformat() if d.updated_at else None,
                 "source": d.source,

hy_memory/coding/extractor.py

@@ -1166,7 +1166,10 @@ class CodingMemoryExtractor:
                     obj = None
 
         if not isinstance(obj, dict):
-            logger.warning(f"[coding-extract] failed to parse output: {text!r}")
+            logger.warning(
+                f"[coding-extract] failed to parse output ({len(text or '')} chars)"
+            )
+            logger.debug(f"[coding-extract] unparsable output: {text!r}")
             return [], []
 
         # "memories" 是双层化之前的 key，老 prompt 缓存 / 旧模型可能还这么回

hy_memory/coding/fact_reconciler.py

@@ -454,9 +454,10 @@ class WorkingFactReconciler:
         workspace_id: Optional[str] = None,
         branch: Optional[str] = None,
         source: str = "default",
+        memory_at: Optional[datetime] = None,
     ) -> Tuple[List[WorkingFact], List[FactReconcileOp]]:
         """
-        执行 reconcile + 持久化。
+        执行 reconcile + 持久化。memory_at 是这段对话的时间，记到新事实上。
 
         Returns:
             (真正落库的新事实, 执行过的 ops)。LLM 挂掉时退化为全 INSERT 不覆盖，
@@ -493,6 +494,7 @@ class WorkingFactReconciler:
             workspace_id=workspace_id,
             branch=branch,
             source=source,
+            memory_at=memory_at,
         )
         logger.info(
             f"[fact-reconcile] drafts={len(drafts)} inserted={len(facts)} "
@@ -568,9 +570,12 @@ class WorkingFactReconciler:
                     kept.append(fact)
             if blocked:
                 logger.info(
-                    f"[fact-reconcile] subject gate blocked {len(blocked)} candidate(s) "
-                    f"for subject={draft.subject!r}: "
-                    + ", ".join(f"{f.fact_id}({f.subject!r})" for f in blocked)
+                    f"[fact-reconcile] subject gate blocked {len(blocked)} candidate(s): "
+                    + ", ".join(f.fact_id for f in blocked)
+                )
+                logger.debug(
+                    f"[fact-reconcile] subject gate: draft subject={draft.subject!r} "
+                    f"blocked subjects={[f.subject for f in blocked]!r}"
                 )
             result.append(kept)
         return result
@@ -743,7 +748,10 @@ class WorkingFactReconciler:
                     obj = None
 
         if not isinstance(obj, dict):
-            logger.warning(f"[fact-reconcile] failed to parse ops output: {text!r}")
+            logger.warning(
+                f"[fact-reconcile] failed to parse ops output ({len(text or '')} chars)"
+            )
+            logger.debug(f"[fact-reconcile] unparsable ops output: {text!r}")
             return {}
         ops_raw = obj.get("ops")
         if not isinstance(ops_raw, list):
@@ -781,6 +789,7 @@ class WorkingFactReconciler:
         workspace_id: Optional[str],
         branch: Optional[str] = None,
         source: str = "default",
+        memory_at: Optional[datetime] = None,
     ) -> Tuple[List[WorkingFact], List[FactReconcileOp]]:
         inserted: List[WorkingFact] = []
         ops: List[FactReconcileOp] = []
@@ -801,6 +810,7 @@ class WorkingFactReconciler:
                 subject=draft.subject,
                 details=draft.details,
                 status="active",
+                memory_at=memory_at,
             )
 
             allowed = {c.fact_id for c in (candidates[idx] if idx < len(candidates) else [])}
@@ -831,8 +841,11 @@ class WorkingFactReconciler:
                 # 别的主体的状态不是。闸门关掉时这道复查一并跳过，否则闸门就没关干净。
                 if self.subject_gate and _subject_conflict(draft.subject, old.subject):
                     logger.warning(
-                        f"[fact-reconcile] refused cross-subject SUPERSEDE: "
-                        f"{old.fact_id}({old.subject!r}) by draft subject {draft.subject!r}"
+                        f"[fact-reconcile] refused cross-subject SUPERSEDE: {old.fact_id}"
+                    )
+                    logger.debug(
+                        f"[fact-reconcile] cross-subject: old subject={old.subject!r} "
+                        f"draft subject={draft.subject!r}"
                     )
                     continue
                 try:

hy_memory/coding/judge.py

@@ -195,8 +195,10 @@ async def classify_messages_scene(
         result = _parse_scene_json(resp.content)
         if result is None:
             logger.warning(
-                f"[coding-judge] failed to parse classify_messages output: {resp.content!r}; defaulting to chat"
+                f"[coding-judge] failed to parse classify_messages output "
+                f"({len(resp.content or '')} chars); defaulting to chat"
             )
+            logger.debug(f"[coding-judge] unparsable classify_messages output: {resp.content!r}")
             await _store_coding_log(
                 cache, request_id=request_id, user_id=user_id, agent_id=agent_id,
                 session_id=session_id, step="CODING_JUDGE", prompt=prompt,
@@ -207,9 +209,8 @@ async def classify_messages_scene(
             return "chat"
         scene = result["scene"]
         parsed = {**result, "is_coding": scene != "chat"}
-        logger.info(
-            f"[coding-judge] scene={scene} reason={result.get('reason', '')!r}"
-        )
+        logger.info(f"[coding-judge] scene={scene}")
+        logger.debug(f"[coding-judge] scene={scene} reason={result.get('reason', '')!r}")
         await _store_coding_log(
             cache, request_id=request_id, user_id=user_id, agent_id=agent_id,
             session_id=session_id, step="CODING_JUDGE", prompt=prompt,
@@ -295,8 +296,10 @@ async def classify_queries_scene(
         result = _parse_scene_json(resp.content)
         if result is None:
             logger.warning(
-                f"[coding-judge] failed to parse classify_queries output: {resp.content!r}; defaulting to chat"
+                f"[coding-judge] failed to parse classify_queries output "
+                f"({len(resp.content or '')} chars); defaulting to chat"
             )
+            logger.debug(f"[coding-judge] unparsable classify_queries output: {resp.content!r}")
             return "chat"
         scene = result["scene"]
         logger.info(f"[coding-judge] queries scene={scene}")
@@ -422,15 +425,18 @@ async def classify_and_rewrite_queries(
     scene = _parse_scene_label(head)
     if scene is None:
         logger.warning(
-            f"[coding-judge] classify_and_rewrite unrecognized head={head!r}; falling back to chat"
+            f"[coding-judge] classify_and_rewrite unrecognized head "
+            f"({len(head)} chars); falling back to chat"
         )
+        logger.debug(f"[coding-judge] classify_and_rewrite unrecognized head={head!r}")
         return {"scene": "chat", "is_coding": False, "rewrite_query": rewrite, "ok": False}
     is_coding = scene != "chat"
 
     logger.info(
         f"[coding-judge] classify_and_rewrite scene={scene} "
-        f"rewrite={rewrite!r}"
+        f"rewrite({len(rewrite)} chars)"
     )
+    logger.debug(f"[coding-judge] classify_and_rewrite rewrite={rewrite!r}")
     return {"scene": scene, "is_coding": is_coding, "rewrite_query": rewrite, "ok": True}
 
 
@@ -569,8 +575,9 @@ async def analyze_queries(
         parsed_scene = _parse_scene_label(head)
         if parsed_scene is None:
             logger.warning(
-                f"[coding-judge] analyze_queries unrecognized head={head!r}; ok=False"
+                f"[coding-judge] analyze_queries unrecognized head ({len(head)} chars); ok=False"
             )
+            logger.debug(f"[coding-judge] analyze_queries unrecognized head={head!r}")
             return {**fallback, "since_raw": ""}
         scene = parsed_scene
         idx = 1
@@ -589,9 +596,10 @@ async def analyze_queries(
     created_after = _since_to_created_after(since_raw, current_time)
 
     logger.info(
-        f"[coding-judge] analyze scene={scene} rewrite={rewrite!r} "
+        f"[coding-judge] analyze scene={scene} rewrite({len(rewrite)} chars) "
         f"since={since_raw!r} created_after={created_after}"
     )
+    logger.debug(f"[coding-judge] analyze rewrite={rewrite!r}")
     return {
         "scene": scene, "is_coding": scene != "chat", "rewrite_query": rewrite,
         "created_after": created_after, "since_raw": since_raw, "ok": True,
@@ -614,7 +622,10 @@ def _since_to_created_after(since_raw: str, current_time: str) -> Optional[float
             return dt.timestamp()
         except (ValueError, TypeError):
             continue
-    logger.warning(f"[coding-judge] analyze_queries unparsable SINCE={since_raw!r}")
+    logger.warning(
+        f"[coding-judge] analyze_queries unparsable SINCE ({len(since_raw)} chars)"
+    )
+    logger.debug(f"[coding-judge] analyze_queries unparsable SINCE={since_raw!r}")
     return None
 
 

hy_memory/coding/preproc.py

@@ -11,6 +11,7 @@ Coding Memory 预处理工具（规则，零 LLM）
 使用方式与 scene 路由见 docs/scene_coding_usage.md。
 """
 
+import re
 from typing import Any, Dict, List, Optional, Tuple
 
 from ..pipelines.base import ChatMessage
@@ -63,6 +64,13 @@ DEFAULT_MAX_TOOL_RESULT_BYTES = 2048
 DEFAULT_HEAD_BYTES = 1024
 DEFAULT_TAIL_BYTES = 512
 
+# 本模块自己产出的省略标记。截断结果会再次经过截断（落库前一次、读历史时又一次），
+# 认出自己的标记才能保证幂等：已截断的文本再截一次原样返回，而不是把省略计数改写。
+_TOOL_OMIT_MARKER_RE = re.compile(r"\n\n\[\.\.\.\d+ bytes omitted by truncate\.\.\.\]\n\n")
+# 标记本身的字节数上限（数字位数有限），用于判断「已截断且只超出一个标记的长度」。
+_TOOL_OMIT_MARKER_MAX_BYTES = 128
+_ASSISTANT_OMIT_MARKER_RE = re.compile(r" \[\.\.\.\d+ chars omitted\.\.\.\] ")
+
 
 def truncate_tool_result_text(
     text: str,
@@ -81,6 +89,12 @@ def truncate_tool_result_text(
     raw = text.encode("utf-8")
     if len(raw) <= max_bytes:
         return text
+    # 已截断过（max_bytes 很小时标记本身会让结果略超上限）：原样返回，保证幂等。
+    if (
+        len(raw) <= max_bytes + _TOOL_OMIT_MARKER_MAX_BYTES
+        and _TOOL_OMIT_MARKER_RE.search(text)
+    ):
+        return text
     # 自定义 max_bytes 可能小于默认 head+tail（1024+512）。按原比例缩放内容窗口，
     # 给省略标记预留空间，不能因为窗口和过大而把超限原文直接返回。
     marker_budget = min(96, max(24, max_bytes // 4))
@@ -161,6 +175,10 @@ def truncate_assistant_text(
         text = "" if text is None else str(text)
     if len(text) <= head_chars + tail_chars:
         return text
+    # 已截断过：中段恰好是本函数的省略标记 → 原样返回，保证幂等（否则会把
+    # 省略计数改写成标记自身的长度）。
+    if _ASSISTANT_OMIT_MARKER_RE.fullmatch(text[head_chars:len(text) - tail_chars]):
+        return text
     omitted = len(text) - head_chars - tail_chars
     return (
         text[:head_chars]

hy_memory/coding/reconciler.py

@@ -388,7 +388,10 @@ class CodingMemoryReconciler:
                     obj = None
 
         if not obj or not isinstance(obj, dict):
-            logger.warning(f"[coding-reconcile] failed to parse ops output: {text!r}")
+            logger.warning(
+                f"[coding-reconcile] failed to parse ops output ({len(text or '')} chars)"
+            )
+            logger.debug(f"[coding-reconcile] unparsable ops output: {text!r}")
             return []
         ops_raw = obj.get("ops") or []
         if not isinstance(ops_raw, list):
@@ -441,10 +444,10 @@ class CodingMemoryReconciler:
         for op in ops:
             if op.action == "ADD":
                 if op.draft_idx is None or not (0 <= op.draft_idx < len(drafts)):
-                    logger.warning(f"[coding-reconcile] drop invalid ADD: {op.to_dict()}")
+                    logger.warning(f"[coding-reconcile] drop invalid ADD: {_op_ref(op)}")
                     continue
                 if op.draft_idx in consumed:
-                    logger.warning(f"[coding-reconcile] drop ADD on consumed draft: {op.to_dict()}")
+                    logger.warning(f"[coding-reconcile] drop ADD on consumed draft: {_op_ref(op)}")
                     continue
                 d = drafts[op.draft_idx]
                 memory = self._draft_to_new_memory(d)
@@ -458,18 +461,18 @@ class CodingMemoryReconciler:
 
             elif op.action == "UPDATE":
                 if op.draft_idx is None or not (0 <= op.draft_idx < len(drafts)):
-                    logger.warning(f"[coding-reconcile] drop invalid UPDATE: {op.to_dict()}")
+                    logger.warning(f"[coding-reconcile] drop invalid UPDATE: {_op_ref(op)}")
                     continue
                 if not op.target_experience_id:
-                    logger.warning(f"[coding-reconcile] drop UPDATE missing target: {op.to_dict()}")
+                    logger.warning(f"[coding-reconcile] drop UPDATE missing target: {_op_ref(op)}")
                     continue
                 if not self._target_was_offered(op, candidates_per_draft):
                     logger.warning(
-                        f"[coding-reconcile] drop UPDATE outside authority scope: {op.to_dict()}"
+                        f"[coding-reconcile] drop UPDATE outside authority scope: {_op_ref(op)}"
                     )
                     continue
                 if op.draft_idx in consumed:
-                    logger.warning(f"[coding-reconcile] drop UPDATE on consumed draft: {op.to_dict()}")
+                    logger.warning(f"[coding-reconcile] drop UPDATE on consumed draft: {_op_ref(op)}")
                     continue
                 d = drafts[op.draft_idx]
                 memory = self._draft_to_new_memory(d, experience_id=op.target_experience_id)
@@ -488,17 +491,17 @@ class CodingMemoryReconciler:
 
             elif op.action == "DELETE":
                 if not op.target_experience_id:
-                    logger.warning(f"[coding-reconcile] drop DELETE missing target: {op.to_dict()}")
+                    logger.warning(f"[coding-reconcile] drop DELETE missing target: {_op_ref(op)}")
                     continue
                 if not self._target_was_offered(op, candidates_per_draft):
                     logger.warning(
-                        f"[coding-reconcile] drop DELETE outside authority scope: {op.to_dict()}"
+                        f"[coding-reconcile] drop DELETE outside authority scope: {_op_ref(op)}"
                     )
                     continue
                 # DELETE 需要知道 user_id；通过 draft_idx（如有）或第一个 draft 推断
                 user_id = self._infer_user_id(op, drafts)
                 if not user_id:
-                    logger.warning(f"[coding-reconcile] drop DELETE no user_id context: {op.to_dict()}")
+                    logger.warning(f"[coding-reconcile] drop DELETE no user_id context: {_op_ref(op)}")
                     continue
                 try:
                     ok = await self.store.delete(op.target_experience_id, user_id=user_id)
@@ -577,6 +580,7 @@ class CodingMemoryReconciler:
             confidence=d.confidence,
             source=d.source,
             extraction_source=d.extraction_source,
+            memory_at=d.memory_at,
         )
 
 
@@ -584,6 +588,14 @@ class CodingMemoryReconciler:
 # Helpers
 # ================================================================
 
+def _op_ref(op: ReconcileOp) -> str:
+    """op 的日志引用：只带动作与编号，不带 LLM 写的 reason。"""
+    return (
+        f"action={op.action} draft_idx={op.draft_idx} "
+        f"target={op.target_experience_id}"
+    )
+
+
 def _truncate(s: str, n: int) -> str:
     if not s:
         return ""

hy_memory/coding/store.py

@@ -86,6 +86,7 @@ CREATE TABLE IF NOT EXISTS coding_memory_meta (
     confidence       REAL NOT NULL DEFAULT 0.7,
     source           TEXT NOT NULL DEFAULT 'default',
     extraction_source TEXT NOT NULL DEFAULT 'auto_extract',
+    memory_at        INTEGER,
     created_at       TEXT NOT NULL,
     updated_at       TEXT NOT NULL
 );
@@ -109,6 +110,7 @@ CREATE TABLE IF NOT EXISTS working_fact (
     derived_experience_ids TEXT NOT NULL DEFAULT '[]',
     status                 TEXT NOT NULL DEFAULT 'active',
     fact_chain_summary     TEXT NOT NULL DEFAULT '',
+    memory_at              INTEGER,
     created_at             TEXT NOT NULL
 );
 -- 反查前驱（谁的 fact_link 指向我）走这个索引，写演变总结时要顺着链往回走
@@ -164,6 +166,8 @@ _ADDED_COLUMNS = {
     "coding_memory_meta": {
         "source_fact_ids": "TEXT NOT NULL DEFAULT '[]'",
         "extraction_source": "TEXT NOT NULL DEFAULT 'auto_extract'",
+        # 对话时间，秒级 epoch（同 chat）。老库里的行是 NULL：那时没存过。
+        "memory_at": "INTEGER",
     },
     # 早期版本存过一列 fact_chain（机械累积的历代说法），已废弃：同样的信息顺着
     # fact_link 反查就有。老库里那一列留着不管，有默认值不影响插入。
@@ -177,6 +181,7 @@ _ADDED_COLUMNS = {
         "subject": "TEXT NOT NULL DEFAULT ''",
         "details": "TEXT NOT NULL DEFAULT ''",
         "source": "TEXT NOT NULL DEFAULT 'default'",
+        "memory_at": "INTEGER",
     },
 }
 
@@ -290,6 +295,29 @@ def _parse_dt(s: Optional[str]) -> Optional[datetime]:
         return None
 
 
+def _from_epoch(value: Any) -> Optional[datetime]:
+    """memory_at 列（秒级 epoch）→ datetime；NULL 或坏值返回 None。"""
+    if value is None:
+        return None
+    try:
+        return datetime.fromtimestamp(int(value))
+    except Exception:
+        return None
+
+
+def _to_epoch(value: Optional[datetime]) -> Optional[int]:
+    return int(value.timestamp()) if isinstance(value, datetime) else None
+
+
+def _later(a: Optional[datetime], b: Optional[datetime]) -> Optional[datetime]:
+    """两个对话时间取较晚的；一边缺就取另一边。"""
+    if a is None:
+        return b
+    if b is None:
+        return a
+    return a if _to_epoch(a) >= _to_epoch(b) else b
+
+
 def _row_to_memory(row: sqlite3.Row) -> CodingMemory:
     """SQLite row → CodingMemory"""
     return CodingMemory(
@@ -308,6 +336,7 @@ def _row_to_memory(row: sqlite3.Row) -> CodingMemory:
         confidence=row["confidence"],
         source=row["source"],
         extraction_source=row["extraction_source"],
+        memory_at=_from_epoch(row["memory_at"]),
         created_at=_parse_dt(row["created_at"]),
         updated_at=_parse_dt(row["updated_at"]),
     )
@@ -334,6 +363,7 @@ def _row_to_fact(row: sqlite3.Row) -> WorkingFact:
         # 老库刚 ADD COLUMN 时可能取不到这一列，用 keys() 兜一下
         fact_chain_summary=(row["fact_chain_summary"] or ""
                             if "fact_chain_summary" in row.keys() else ""),
+        memory_at=_from_epoch(row["memory_at"]),
         created_at=_parse_dt(row["created_at"]),
     )
 
@@ -575,7 +605,11 @@ class CodingMemoryStore:
         await self._upsert_vec_keys(memory)
         logger.info(
             f"[coding-store] inserted experience_id={memory.experience_id} "
-            f"task={memory.task!r} keys={1 + len(memory.search_keys)}"
+            f"keys={1 + len(memory.search_keys)}"
+        )
+        logger.debug(
+            f"[coding-store] inserted experience_id={memory.experience_id} "
+            f"task={memory.task!r}"
         )
 
     async def update(self, memory: CodingMemory) -> None:
@@ -597,6 +631,8 @@ class CodingMemoryStore:
         memory.updated_at = datetime.utcnow()
         if memory.created_at is None:
             memory.created_at = prev.created_at
+        # 合并后的经验讲的是两段对话，对话时间取较晚的那段
+        memory.memory_at = _later(prev.memory_at, memory.memory_at)
         await _run_in_pool(self._upsert_sql_sync, memory)
         # 删旧 VDB
         await self._delete_vec_keys_by_count(memory.experience_id, prev_key_count)
@@ -970,7 +1006,8 @@ class CodingMemoryStore:
             fact.created_at = datetime.utcnow()
         await _run_in_pool(self._upsert_fact_sql_sync, fact)
         await self._upsert_fact_vec(fact)
-        logger.info(
+        logger.info(f"[coding-store] inserted fact_id={fact.fact_id}")
+        logger.debug(
             f"[coding-store] inserted fact_id={fact.fact_id} "
             f"statement={fact.statement!r}"
         )
@@ -1199,8 +1236,8 @@ class CodingMemoryStore:
                     experience_id, user_id, agent_id, task, search_keys, solution,
                     boundary_envs, workspace_id, branch, session_id,
                     files, source_fact_ids, confidence, source, extraction_source,
-                    created_at, updated_at
-                ) VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
+                    memory_at, created_at, updated_at
+                ) VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
                 ON CONFLICT(experience_id) DO UPDATE SET
                     user_id         = excluded.user_id,
                     agent_id        = excluded.agent_id,
@@ -1216,6 +1253,7 @@ class CodingMemoryStore:
                     confidence      = excluded.confidence,
                     source          = excluded.source,
                     extraction_source = excluded.extraction_source,
+                    memory_at       = excluded.memory_at,
                     updated_at      = excluded.updated_at
                 """,
                 (
@@ -1226,6 +1264,7 @@ class CodingMemoryStore:
                     json.dumps(memory.files, ensure_ascii=False),
                     json.dumps(memory.source_fact_ids, ensure_ascii=False),
                     memory.confidence, memory.source, memory.extraction_source,
+                    _to_epoch(memory.memory_at),
                     (memory.created_at or datetime.utcnow()).isoformat(),
                     (memory.updated_at or datetime.utcnow()).isoformat(),
                 ),
@@ -1366,8 +1405,8 @@ class CodingMemoryStore:
                     fact_id, user_id, agent_id, session_id, workspace_id, branch,
                     source, statement, files, subject, details,
                     fact_link, derived_experience_ids,
-                    status, fact_chain_summary, created_at
-                ) VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
+                    status, fact_chain_summary, memory_at, created_at
+                ) VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
                 ON CONFLICT(fact_id) DO UPDATE SET
                     user_id                = excluded.user_id,
                     agent_id               = excluded.agent_id,
@@ -1382,7 +1421,8 @@ class CodingMemoryStore:
                     fact_link              = excluded.fact_link,
                     derived_experience_ids = excluded.derived_experience_ids,
                     status                 = excluded.status,
-                    fact_chain_summary     = excluded.fact_chain_summary
+                    fact_chain_summary     = excluded.fact_chain_summary,
+                    memory_at              = excluded.memory_at
                 """,
                 (
                     fact.fact_id, fact.user_id, fact.agent_id,
@@ -1395,6 +1435,7 @@ class CodingMemoryStore:
                     json.dumps(fact.derived_experience_ids, ensure_ascii=False),
                     fact.status,
                     fact.fact_chain_summary or "",
+                    _to_epoch(fact.memory_at),
                     (fact.created_at or datetime.utcnow()).isoformat(),
                 ),
             )

hy_memory/coding/types.py

@@ -25,6 +25,11 @@ FactStatus = Literal["active", "superseded"]
 MAX_FACT_CHAIN_ENTRIES = 8
 
 
+def _epoch(value: Optional[datetime]) -> Optional[int]:
+    """memory_at 对外是秒级 epoch，与 chat 一致。"""
+    return int(value.timestamp()) if isinstance(value, datetime) else None
+
+
 # ================================================================
 # 草稿（LLM extract 阶段输出）
 # ================================================================
@@ -53,6 +58,8 @@ class CodingMemoryDraft:
     files: List[str] = field(default_factory=list)
     source: str = "default"
     extraction_source: str = "auto_extract"
+    # 对话时间（SDK 注入，同 chat 的 memory_at）
+    memory_at: Optional[datetime] = None
 
     def __post_init__(self) -> None:
         from ..utils.source_filter import normalize_source
@@ -73,6 +80,7 @@ class CodingMemoryDraft:
             "files": list(self.files),
             "source": self.source,
             "extraction_source": self.extraction_source,
+            "memory_at": _epoch(self.memory_at),
         }
 
 
@@ -111,6 +119,9 @@ class CodingMemory:
     extraction_source: str = "auto_extract"
 
     # 时间
+    # memory_at 是对话时间（同 chat），created_at / updated_at 是写库时间。
+    # UPDATE 合并时取两者中较晚的那个对话时间，见 store.update。
+    memory_at: Optional[datetime] = None
     created_at: Optional[datetime] = None
     updated_at: Optional[datetime] = None
 
@@ -135,6 +146,7 @@ class CodingMemory:
             "confidence": self.confidence,
             "source": self.source,
             "extraction_source": self.extraction_source,
+            "memory_at": _epoch(self.memory_at),
             "created_at": self.created_at.isoformat() if self.created_at else None,
             "updated_at": self.updated_at.isoformat() if self.updated_at else None,
         }
@@ -249,6 +261,7 @@ class WorkingFact:
     fact_chain_summary: str = ""
 
     # 时间
+    memory_at: Optional[datetime] = None   # 对话时间（同 chat），没传时是收到请求的时间
     created_at: Optional[datetime] = None  # 产生时间，新的优先
 
     def __post_init__(self) -> None:
@@ -272,6 +285,7 @@ class WorkingFact:
             "derived_experience_ids": list(self.derived_experience_ids),
             "status": self.status,
             "fact_chain_summary": self.fact_chain_summary,
+            "memory_at": _epoch(self.memory_at),
             "created_at": self.created_at.isoformat() if self.created_at else None,
         }
 

hy_memory/coding/writer.py

@@ -18,6 +18,7 @@ CodingWriter - 写入端协调器
 import logging
 import time
 import uuid
+from datetime import datetime
 from typing import Any, Dict, List, Optional
 
 from ..agent.llm_provider import LLMProvider
@@ -79,9 +80,11 @@ class CodingWriter:
         session_id: Optional[str] = None,
         request_id: Optional[str] = None,
         source: str = "default",
+        memory_at: Optional[datetime] = None,
     ) -> Dict[str, Any]:
         """
-        working/coding 段写入端到端。返回响应 dict（兼容 client.add 返回结构）：
+        working/coding 段写入端到端。memory_at 是对话时间，记到这次写出的经验和
+        事实上；不传则用收到请求的时间（与 chat 链路一致）。返回响应 dict（兼容 client.add 返回结构）：
             {
                 "success": True,
                 "scene": "productivity",
@@ -101,6 +104,8 @@ class CodingWriter:
         """
         request_id = request_id or str(uuid.uuid4())
         t0 = time.perf_counter()
+        if memory_at is None:
+            memory_at = datetime.now()
 
         if not messages:
             return self._empty_response(request_id, t0)
@@ -172,6 +177,7 @@ class CodingWriter:
         # extractor 只负责内容；调用来源由 SDK 入口注入，不能交给 LLM 生成。
         for draft in drafts:
             draft.source = source
+            draft.memory_at = memory_at
 
         # 3. 经验层 reconcile + 持久化
         ops: List[ReconcileOp] = (
@@ -197,6 +203,7 @@ class CodingWriter:
                 workspace_id=workspace_id,
                 branch=branch,
                 source=source,
+                memory_at=memory_at,
             )
 
         experience_ids = [op.target_experience_id for op in ops

hy_memory/config.py

@@ -260,6 +260,7 @@ class EmbedderConfig:
     max_retries: int = None  # 最大重试次数
     retry_delay: float = None  # 重试间隔（秒）
     timeout: int = None  # 请求超时（秒）
+    batch_size: int = None  # 单次 HTTP 请求最多带多少条 input（embed_batch 按此切块）
 
     # 内部平台扩展（extra_headers / extra_body 会透传给 OpenAI client）
     extra_headers: Optional[Dict[str, str]] = None
@@ -294,6 +295,10 @@ class EmbedderConfig:
             # 请求切断并制造重试风暴。SDK 内层重试已关闭，15s 是每个显式 attempt
             # 的真实上限；部署可通过 MEMORY_EMBEDDER_TIMEOUT 覆盖。
             self.timeout = _get_env_int("MEMORY_EMBEDDER_TIMEOUT", 15)
+        if self.batch_size is None:
+            # 大批量（如重新 embed 整个库）不能一次 HTTP 发几千条：多数服务对
+            # input 数组有上限，超了直接 400。按 batch_size 切块逐块发送。
+            self.batch_size = _get_env_int("MEMORY_EMBEDDER_BATCH_SIZE", 64)
         if self.extra_headers is None:
             val = os.getenv("MEMORY_EMBEDDER_EXTRA_HEADERS")
             if val:

hy_memory/core/embed_service.py

@@ -38,10 +38,12 @@ import importlib.util
 import json
 from json import JSONDecodeError
 import logging
+import random
 import threading
 import time
 
 from ..config import MemoryConfig
+from ..embedding import EmbeddingDimensionMismatchError
 from .workload import embedding_slot
 from ..utils.log_setup import get_request_id
 
@@ -80,7 +82,7 @@ def _count_tokens(text: str) -> int:
     try:
         import tiktoken
 
-        return len(tiktoken.get_encoding("cl100k_base").encode(text))
+        return len(tiktoken.get_encoding("cl100k_base").encode(text, disallowed_special=()))
     except Exception:
         return len(text) // _CHARS_PER_TOKEN_FALLBACK + 1
 
@@ -98,7 +100,8 @@ def _truncate_for_embedding(text: str, limit: int = _EMBED_MAX_TOKENS) -> str:
         import tiktoken
 
         enc = tiktoken.get_encoding("cl100k_base")
-        ids = enc.encode(text)
+        # Input is ordinary user text, not a tokenizer control-token stream.
+        ids = enc.encode(text, disallowed_special=())
         if len(ids) <= limit:
             return text
         dropped = len(ids) - limit
@@ -186,6 +189,41 @@ def _redact_headers(headers: Optional[Dict[str, Any]]) -> Optional[Dict[str, Any
     return out
 
 
+# Retry-After 的上限：服务端给出超长等待（或恶意值）时不让单次调用挂几十分钟。
+_RETRY_AFTER_MAX_S = 60.0
+
+
+def _retry_after_seconds(exc: Exception) -> Optional[float]:
+    """从 429 异常的响应头里取 Retry-After（秒数或 HTTP 日期），取不到返回 None。"""
+    response = getattr(exc, "response", None)
+    headers = getattr(response, "headers", None)
+    if not headers:
+        return None
+    try:
+        raw = headers.get("retry-after") or headers.get("Retry-After")
+    except Exception:
+        return None
+    if raw is None:
+        return None
+    raw = str(raw).strip()
+    try:
+        seconds = float(raw)
+    except ValueError:
+        try:
+            from email.utils import parsedate_to_datetime
+            from datetime import datetime, timezone
+
+            when = parsedate_to_datetime(raw)
+            if when.tzinfo is None:
+                when = when.replace(tzinfo=timezone.utc)
+            seconds = (when - datetime.now(timezone.utc)).total_seconds()
+        except Exception:
+            return None
+    if seconds != seconds:  # NaN
+        return None
+    return min(max(0.0, seconds), _RETRY_AFTER_MAX_S)
+
+
 def _embeddings_api_url(base_url: str) -> str:
     """OpenAI SDK posts to ``{base_url}/embeddings`` (base_url usually ends with /v1)."""
     base = (base_url or "").rstrip("/")
@@ -204,7 +242,8 @@ class EmbedServiceConfig:
     api_key: str = ""
     base_url: str = ""
     embedding_dims: int = 1536
-    batch_size: int = 100
+    # embed_batch 单次 HTTP 最多带多少条 input；超出按块顺序发送
+    batch_size: int = 64
     enable_cache: bool = True
     cache_size: int = 10000
     max_retries: int = 5
@@ -286,6 +325,7 @@ class EmbedService:
             max_retries=embedder.max_retries if embedder.max_retries is not None else 5,
             retry_delay=embedder.retry_delay if embedder.retry_delay is not None else 3.0,
             timeout=embedder.timeout if embedder.timeout is not None else 60,
+            batch_size=max(1, int(getattr(embedder, "batch_size", None) or 64)),
             queue_batch_size=int(os.getenv("MEMORY_EMBED_BATCH_SIZE", "8")),
             queue_batch_window_ms=float(os.getenv("MEMORY_EMBED_BATCH_WINDOW_MS", "100")),
             extra_headers=embedder.extra_headers,
@@ -639,6 +679,8 @@ class EmbedService:
         self,
         texts: List[str],
         use_cache: bool = True,
+        *,
+        max_retries: Optional[int] = None,
     ) -> List[List[float]]:
         """
         批量向量化
@@ -646,17 +688,23 @@ class EmbedService:
         Args:
             texts: 文本列表
             use_cache: 是否使用缓存
+            max_retries: 本次调用的重试次数；None 用配置值。长批量任务（重新
+                embed 整个库）可以给更宽的重试预算，不影响在线路径。
         
         Returns:
             向量列表
         """
-        vectors, _metas = await self.embed_batch_with_meta(texts, use_cache=use_cache)
+        vectors, _metas = await self.embed_batch_with_meta(
+            texts, use_cache=use_cache, max_retries=max_retries,
+        )
         return vectors
 
     async def embed_batch_with_meta(
         self,
         texts: List[str],
         use_cache: bool = True,
+        *,
+        max_retries: Optional[int] = None,
     ) -> Tuple[List[List[float]], List[Dict[str, Any]]]:
         """批量向量化，同时返回每条的 provider metadata（非 vector）。"""
         if not texts:
@@ -681,7 +729,12 @@ class EmbedService:
             indices_to_embed.append(i)
 
         if texts_to_embed:
-            vectors, batch_metas = await self._embed_texts_with_meta(texts_to_embed)
+            # max_retries 只在显式给出时透传：保持 `_embed_texts_with_meta(texts)`
+            # 这个既有调用形状不变（测试与下游替身按它打桩）。
+            retry_kwargs = {} if max_retries is None else {"max_retries": max_retries}
+            vectors, batch_metas = await self._embed_texts_with_meta(
+                texts_to_embed, **retry_kwargs,
+            )
 
             for j, vector in enumerate(vectors):
                 idx = indices_to_embed[j]
@@ -699,9 +752,9 @@ class EmbedService:
         return results, metas
     
     async def _embed_texts_with_meta(
-        self, texts: List[str]
+        self, texts: List[str], *, max_retries: Optional[int] = None,
     ) -> Tuple[List[List[float]], List[Dict[str, Any]]]:
-        """实际的向量化调用：**一次 HTTP 带走整批**；返回 (vectors, metas)。
+        """实际的向量化调用：**一次 HTTP 带走一块**；返回 (vectors, metas)。
 
         早期实现是 `for text in texts: await _embed_openai([text])` 逐条串行，
         原意是"避免 batch size 限制"，但代价极大：线上一批 8 条 × ~5.6s 串行
@@ -718,21 +771,34 @@ class EmbedService:
         截断在这里做（发送前逐条过 `_truncate_for_embedding`）：所有 embedding
         路径（embed / embed_batch / embed_queued / *_with_meta）最终都落到这个
         方法 —— 放在各入口会漏。理由与阈值见模块头部 `_EMBED_MAX_TOKENS`。
+
+        切块：一次带走整批只适用于小批。几千条的 input 数组会被多数服务直接
+        400（input 上限），所以按 `batch_size`（默认 64）切块，块之间顺序发送。
+        不超过一块时行为与切块前逐字相同（`batch_mode="single_request"`）。
         """
         if not texts:
             return [], []
 
         safe_texts = [_truncate_for_embedding(t) for t in texts]
-        vectors, meta = await self._embed_openai(safe_texts)
-        base_meta = dict(meta or {})
-        base_meta["batch_mode"] = "single_request"
-        base_meta["batch_input_count"] = len(safe_texts)
-
+        chunk = max(1, int(self._embed_config.batch_size or 64))
+        vectors: List[List[float]] = []
         all_metas: List[Dict[str, Any]] = []
-        for idx in range(len(safe_texts)):
-            m = dict(base_meta)
-            m["batch_index"] = idx
-            all_metas.append(m)
+        chunk_count = (len(safe_texts) + chunk - 1) // chunk
+        for start in range(0, len(safe_texts), chunk):
+            part = safe_texts[start:start + chunk]
+            retry_kwargs = {} if max_retries is None else {"max_retries": max_retries}
+            part_vectors, meta = await self._embed_openai(part, **retry_kwargs)
+            base_meta = dict(meta or {})
+            base_meta["batch_mode"] = "single_request" if chunk_count == 1 else "chunked"
+            base_meta["batch_input_count"] = len(part)
+            if chunk_count > 1:
+                base_meta["batch_chunk_index"] = start // chunk
+                base_meta["batch_chunk_count"] = chunk_count
+            for idx in range(len(part)):
+                m = dict(base_meta)
+                m["batch_index"] = start + idx
+                all_metas.append(m)
+            vectors.extend(part_vectors)
 
         self._last_embed_meta = all_metas[-1] if all_metas else {}
         return vectors, all_metas
@@ -753,16 +819,22 @@ class EmbedService:
         """
         打一条 INFO 级 embedding 请求日志（含 request_id / provider / model / api / 入参 / 出参）。
 
-        出参只含 vector 维度，不含向量数值。request_id 同时写入 message，
+        出参只含 vector 维度，不含向量数值。入参的 input 是待向量化的原文
+        （对话 / 记忆 / 查询），INFO 只记每条的字符数（input_chars），原文仅在
+        DEBUG 单独输出。request_id 同时写入 message，
         并依赖 log formatter 的 %(request_id)s（setup_logging 后生效）。
         """
+        texts = request_payload.get("input")
+        info_request = {k: v for k, v in request_payload.items() if k != "input"}
+        if isinstance(texts, list):
+            info_request["input_chars"] = [len(str(t)) for t in texts]
         payload = {
             "request_id": get_request_id(),
             "status": status,
             "provider": self._embed_config.provider,
             "model": self._embed_config.model,
             "api": api_url,
-            "request": request_payload,
+            "request": info_request,
             "response": response_payload,
         }
         try:
@@ -770,9 +842,13 @@ class EmbedService:
         except Exception:
             body = str(payload)
         logger.info(f"[embed] {body}")
+        if texts is not None and logger.isEnabledFor(logging.DEBUG):
+            logger.debug(
+                f"[embed] input: {json.dumps(texts, ensure_ascii=False, default=str)}"
+            )
 
     async def _embed_openai(
-        self, texts: List[str]
+        self, texts: List[str], *, max_retries: Optional[int] = None,
     ) -> Tuple[List[List[float]], Dict[str, Any]]:
         """
         使用 OpenAI 兼容接口向量化
@@ -862,7 +938,9 @@ class EmbedService:
             "endpoint_overridden": bool(_ov),
         }
 
-        _max_retries = max(0, int(self._embed_config.max_retries or 0))
+        _max_retries = max(0, int(
+            (self._embed_config.max_retries or 0) if max_retries is None else max_retries
+        ))
         _retry_delay = max(0.0, float(self._embed_config.retry_delay or 0.0))
         last_exc: Optional[Exception] = None
 
@@ -908,6 +986,17 @@ class EmbedService:
                         f"got={len(vectors)}"
                     )
                 dims = [len(v) for v in vectors]
+                expected_dims = self._embed_config.embedding_dims
+                if expected_dims:
+                    bad = [d for d in dims if d != int(expected_dims)]
+                    if bad:
+                        # 宽度不对的向量写进集合会把集合的维度钉死在错误值上（Chroma
+                        # 首次 add 定维），或在已有集合上整批失败。确定性错误，不重试。
+                        raise EmbeddingDimensionMismatchError(
+                            "embedding dimension mismatch: "
+                            f"expected={int(expected_dims)} got={bad[0]} "
+                            f"model={self._embed_config.model}"
+                        )
                 meta = {
                     "http_status": raw.status_code,
                     "response_headers": resp_headers,
@@ -943,6 +1032,8 @@ class EmbedService:
                     },
                 )
                 return vectors, meta
+            except EmbeddingDimensionMismatchError:
+                raise
             except Exception as e:
                 last_exc = e
                 _err_name = type(e).__name__
@@ -976,12 +1067,18 @@ class EmbedService:
                 if _retryable and attempt < _max_retries:
                     # 逐出可能卡死的 client,下个 attempt 重建干净连接(恢复原防御语义)
                     self._evict_client(_eff_base_url, _eff_api_key, _eff_timeout)
+                    _delay = _retry_delay
+                    _retry_after = _retry_after_seconds(e) if _status == 429 else None
+                    if _retry_after is not None:
+                        # 限流时服务端明确告诉了要等多久：按它等，再加一点抖动，
+                        # 避免多个 worker 在同一时刻一起醒来再次撞限流。
+                        _delay = max(_delay, _retry_after) + random.uniform(0, 0.25 * _retry_after)
                     logger.warning(
                         f"[embed] attempt {attempt + 1}/{_max_retries + 1} failed "
-                        f"({_err_name}): {_err_msg[:200]}; retry in {_retry_delay}s"
+                        f"({_err_name}): {_err_msg[:200]}; retry in {_delay:.2f}s"
                     )
-                    if _retry_delay > 0:
-                        await asyncio.sleep(_retry_delay)
+                    if _delay > 0:
+                        await asyncio.sleep(_delay)
                     continue
                 self._log_embed_call(
                     status="error",

hy_memory/data/graph_store_kuzu.py

@@ -22,6 +22,7 @@ import logging
 import json
 import threading
 import time
+import weakref
 
 from ..models.memory import MemoryNode, MemoryLayer, MemoryStatus
 from ..config import MemoryConfig
@@ -77,6 +78,21 @@ async def _run_in_kuzu_pool(func, *args, **kwargs):
     return await run_in_executor_guarded(_kuzu_executor, _measured)
 
 
+# 本进程里已打开、尚未 close 的 store。drop_graph 据此拒绝删掉一个正被
+# 活着的 client 使用的图库。WeakSet：store 被回收后自动消失。
+_LIVE_STORES: "weakref.WeakSet[KuzuGraphStore]" = weakref.WeakSet()
+
+
+def open_graph_paths() -> List[str]:
+    """Real paths of the graph databases live stores in this process hold open."""
+    import os
+    return [
+        os.path.realpath(str(store._db_path))
+        for store in list(_LIVE_STORES)
+        if store._db is not None
+    ]
+
+
 def kuzu_executor_snapshot() -> Dict[str, Any]:
     with _kuzu_stats_lock:
         return {
@@ -350,6 +366,7 @@ class KuzuGraphStore(GraphStoreBase):
             )
             self._db = db
             self._conn = conn
+            _LIVE_STORES.add(self)
         except asyncio.TimeoutError:
             logger.warning(
                 f"GraphStore init timed out after 5s (Kuzu file lock contention). "
@@ -2441,6 +2458,7 @@ class KuzuGraphStore(GraphStoreBase):
     async def close(self) -> None:
         """关闭连接"""
         conn, db = self._conn, self._db
+        _LIVE_STORES.discard(self)
         self._conn = None
         self._db = None
         self._available = False

hy_memory/data/vector_store_base.py

@@ -89,6 +89,17 @@ def passes_created_after(
 logger = logging.getLogger(__name__)
 
 
+def physical_collection_name(base_name: Optional[str], dims: Optional[int]) -> str:
+    """The collection name a store opens for ``base_name`` at ``dims``.
+
+    The dims suffix keeps two embedding widths from sharing one collection.
+    `hy_memory.reembed.vector_space_names` uses this too, so callers never
+    re-derive the rule.
+    """
+    base = base_name or "agent_memories"
+    return f"{base}_{dims}" if dims else base
+
+
 class VectorStoreBase(ABC):
     """
     向量存储抽象基类
@@ -98,13 +109,10 @@ class VectorStoreBase(ABC):
 
     def __init__(self, config: MemoryConfig):
         self.config = config
-        base_name = config.vector_store.collection_name or "agent_memories"
-        dims = config.vector_store.embedding_dims
         # 自动拼接维度后缀，避免不同 embedding 模型写入同一 collection 导致维度冲突
-        if dims:
-            self._collection_name = f"{base_name}_{dims}"
-        else:
-            self._collection_name = base_name
+        self._collection_name = physical_collection_name(
+            config.vector_store.collection_name, config.vector_store.embedding_dims,
+        )
 
     _supports_retrieval_clauses: bool = False
     _supports_narrow_episode_queries: bool = False

hy_memory/data/vector_store_chroma.py

@@ -9,10 +9,12 @@ Agent Memory V2 - VectorStore (Chroma)
 - 同步 API → _run_in_vdb_pool() 包装为异步
 """
 
-from typing import Optional, List, Dict, Any
+from typing import Optional, List, Dict, Any, Set, Tuple
 import concurrent.futures
 import functools
 import logging
+import os
+import weakref
 
 from ..models.memory import (
     MemoryNode, MemoryLayer, MemoryStatus,
@@ -33,6 +35,30 @@ _vdb_executor = concurrent.futures.ThreadPoolExecutor(
 )
 
 
+# 本进程里已 initialize、尚未 close 的 store。drop_vector_space 据此拒绝删掉
+# 一个正被活着的 client 使用的集合。WeakSet：store 被回收后自动消失。
+_LIVE_STORES: "weakref.WeakSet[ChromaVectorStore]" = weakref.WeakSet()
+
+
+def _location_key(vs_config) -> Tuple[str, str]:
+    if getattr(vs_config, "host", None):
+        return ("http", f"{vs_config.host}:{vs_config.port or 8000}")
+    return ("local", os.path.realpath(str(vs_config.persist_directory or "")))
+
+
+def open_collection_names(location: Tuple[str, str]) -> Set[str]:
+    """Collections that live stores in this process hold open at ``location``."""
+    names: Set[str] = set()
+    for store in list(_LIVE_STORES):
+        if store._client is None or _location_key(store.config.vector_store) != location:
+            continue
+        names.add(store._collection_name)
+        if store._entity_collection is not None:
+            from ..pipelines._retrieval.config import entity_collection_name
+            names.add(entity_collection_name(store._collection_name))
+    return names
+
+
 async def _run_in_vdb_pool(func, *args, **kwargs):
     """Run a synchronous Chroma call in the dedicated VDB executor."""
     call = functools.partial(func, *args, **kwargs) if args or kwargs else func
@@ -88,6 +114,7 @@ class ChromaVectorStore(VectorStoreBase):
             return client, collection
 
         self._client, self._collection = await _run_in_vdb_pool(_init)
+        _LIVE_STORES.add(self)
         logger.debug(f"VectorStore initialized (Chroma), collection={self._collection_name}")
 
     # ================================================================
@@ -561,6 +588,7 @@ class ChromaVectorStore(VectorStoreBase):
     async def close(self) -> None:
         """关闭客户端"""
         # Chroma PersistentClient 无需显式关闭
+        _LIVE_STORES.discard(self)
         self._client = None
         self._collection = None
         self._entity_collection = None

hy_memory/pipelines/_retrieval/reconcile_retrieval.py

@@ -115,11 +115,16 @@ class ReconcileHybridRetriever:
         fused_max = max(fused_all) if fused_all else 0.0
 
         base = (
-            f"[reconcile-hybrid][{req}] pool={pool_size} q_terms={q_terms!r} "
+            f"[reconcile-hybrid][{req}] pool={pool_size} q_terms={len(q_terms)} "
             f"raw_bm25 max={raw_max:.4f} nonzero={raw_nonzero}/{len(raw_bm25)} "
             f"bm25_norm max={norm_max:.4f} nonzero={norm_nonzero}/{len(bm25_norms)} "
             f"sem_max={sem_max:.4f} fused_max={fused_max:.4f} "
             f"kept={kept}/{len(fused_all)} min_score={self._cfg.min_score} "
+            f"query({len(query_text or '')} chars)"
+        )
+        # query 是待对账的新记忆原文，词项与预览只进 DEBUG
+        logger.debug(
+            f"[reconcile-hybrid][{req}] q_terms={q_terms!r} "
             f"query={q_preview!r} lemmatized={lem_preview!r}"
         )
 

hy_memory/pipelines/cross_domain_sweeper.py

@@ -692,9 +692,10 @@ class CrossDomainSweeper:
                 )
 
             logger.info(
-                f"[sweeper] Created L6 core {core_id[:12]}: '{core_content[:60]}' "
+                f"[sweeper] Created L6 core {core_id[:12]} "
                 f"from {len(cluster_ids)} basics, evidence_inherited={len(inherited_evidence_ids)}"
             )
+            logger.debug(f"[sweeper] L6 core {core_id[:12]} content={core_content[:60]!r}")
             return 1, 0, [core_id]
 
         except Exception as e:

hy_memory/pipelines/reader_base.py

@@ -3587,8 +3587,9 @@ class BaseReadPipeline(ReadPipeline):
                 f"hidden_l0={len(hidden_l0_hits)} fwd_l6={len(forward_l6)} rev_l6={len(reverse_l6)} "
                 f"fused_l6={len(fused_l6)} graph_intent={len(graph_intention_hits)} "
                 f"returned={len(final_results)} "
-                f"query='{request.query[:80]}'"
+                f"query({len(request.query or '')} chars)"
             )
+            logger.debug(f"[read/{self.VERSION}] query={request.query[:80]!r}")
 
             response.success = True
 

hy_memory/pipelines/system2_agent.py

@@ -2516,7 +2516,8 @@ def _parse_operations_json(text: str) -> Optional[List[Dict[str, Any]]]:
             return result
         return None
     except json.JSONDecodeError:
-        logger.warning(f"[S2-agent] JSON parse failed: {json_str[:200]}")
+        logger.warning(f"[S2-agent] JSON parse failed ({len(json_str)} chars)")
+        logger.debug(f"[S2-agent] unparsable output: {json_str[:200]!r}")
         return None
 
 

hy_memory/pipelines/writer.py

@@ -280,6 +280,22 @@ class MemoryWriter(WritePipeline):
             custom["source_raw_memory_ids"] = source_ids
         return custom
 
+    def _flat_trimmed_raw(self, request: WriteRequest) -> str:
+        """把 request.messages 按 raw 统一规则裁剪后拼成 L1_RAW 文本。
+
+        只裁落库副本；request.messages 不动（本次 extractor 输入仍用完整 messages）。
+        """
+        import copy as _copy
+        from ..utils.raw_trim import trim_raw_messages
+
+        _coding_cfg = getattr(getattr(self, "config", None), "coding", None)
+        _trimmed_req = _copy.copy(request)
+        _trimmed_req.messages = trim_raw_messages(
+            request.messages,
+            tool_max_bytes=getattr(_coding_cfg, "tool_result_max_bytes", None),
+        )
+        return _trimmed_req.get_flat_content()
+
     @staticmethod
     def _attach_client_meta(custom: Dict[str, Any], request: Optional[WriteRequest]) -> None:
         """给新建节点挂上调用方 meta（单份 list）。
@@ -2554,9 +2570,14 @@ class MemoryWriter(WritePipeline):
         try:
             # 参数校验
             content = request.content
+            # raw 在写入时就按统一规则裁好（见 utils/raw_trim.py）：工具输出、长
+            # assistant 回复不再原样长期存储；历史拼接直接用存下来的文本，不再二次裁剪。
+            # request.messages 保持原样——本次 extractor 输入仍是完整 messages。
+            _raw_trimmed = False
             if not content and request.has_messages():
-                content = request.get_flat_content()
+                content = self._flat_trimmed_raw(request)
                 request.content = content
+                _raw_trimmed = True
 
             if not content:
                 response.error_code = 400
@@ -2627,6 +2648,9 @@ class MemoryWriter(WritePipeline):
                     vector_store = await self._get_vector_store()
                     layer_enum = MemoryLayer.from_string(suggested_layer)
                     _raw_custom: Dict[str, Any] = {}
+                    if _raw_trimmed:
+                        from ..utils.raw_trim import RAW_TRIMMED_KEY
+                        _raw_custom[RAW_TRIMMED_KEY] = True
                     self._attach_client_meta(_raw_custom, request)
                     mem_node = MemoryNode(
                         user_id=request.user_id,
@@ -2929,13 +2953,16 @@ class MemoryWriter(WritePipeline):
         agent_id: Optional[str] = None,
         exclude_memory_id: str = "",
         max_turns: int = 20,
-        max_chars_assistant: int = 500,
+        max_chars_assistant: Optional[int] = None,
     ) -> str:
         """
         获取该用户最近的 L1_RAW 对话记录，按轮次拆分后拼成 history context。
 
         每条 L1_RAW 的 content 是多轮对话（[user]: ...\n[assistant]: ...）。
-        拆成单条 message 后，user 消息完整保留，assistant 消息截取前 max_chars_assistant 字符。
+        raw 在写入时已按 utils/raw_trim.py 的规则裁好（custom[RAW_TRIMMED_KEY]），
+        这里拆成单条 message 后直接拼接，不再二次裁剪。该规则之前写入的旧 raw
+        没有这个标记，仍按同一（幂等）规则裁剪后再拼，避免旧的长回复 / 工具输出
+        撑大 prompt。max_chars_assistant 只作用于这些旧 raw，默认同写入规则。
         最终取最近 max_turns 条 message（注意单位是 message 条数，非对话轮数；
         调用方按 k 轮 × 2 传入）。
 
@@ -2977,23 +3004,33 @@ class MemoryWriter(WritePipeline):
             # 避免把正文里形如 "[注]: xxx" 的普通文本误判成消息边界。
             import re as _re
             from .base import parse_user_name_line
-            _KNOWN_ROLES = {"user", "assistant", "system", "tool"}
             # 带方括号的任意 label（含对方真人名字）
             _labeled_prefix = _re.compile(r"^\[([^\]\n]{1,32})\]:\s?")
             # 无方括号的已知 role（历史数据容错，沿用旧正则的 `\[?...\]?` 宽容度）
             _bare_role_prefix = _re.compile(r"^(user|assistant|system|tool):\s?")
 
-            # (label, content) 元组列表；label 为 role 或对方真人名字
-            parsed: List[Tuple[str, str]] = []
+            from ..utils.raw_trim import (
+                DEFAULT_RAW_ASSISTANT_MAX_CHARS,
+                RAW_TRIMMED_KEY,
+                trim_raw_body,
+            )
+            if max_chars_assistant is None:
+                max_chars_assistant = DEFAULT_RAW_ASSISTANT_MAX_CHARS
+            _coding_cfg = getattr(getattr(self, "config", None), "coding", None)
+            _tool_max_bytes = getattr(_coding_cfg, "tool_result_max_bytes", None)
+
+            # (label, content, 是否旧 raw 需兼容裁剪) 元组列表；label 为 role 或对方真人名字
+            parsed: List[Tuple[str, str, bool]] = []
             user_name: str = ""
             for node in nodes:
                 content = node.content or ""
+                _legacy = not bool((node.custom or {}).get(RAW_TRIMMED_KEY))
                 for _line_idx, line in enumerate(content.split("\n")):
                     m = _labeled_prefix.match(line) or _bare_role_prefix.match(line)
                     if m:
                         label = m.group(1)
                         body = line[m.end():]
-                        parsed.append((label, body))
+                        parsed.append((label, body, _legacy))
                         continue
                     # 名字行只可能是每个 node content 的第一行（get_flat_content 插在文首）。
                     # 逐 node 判断而非用 `not parsed`：parsed 跨 node 累积，第二个 node
@@ -3006,18 +3043,18 @@ class MemoryWriter(WritePipeline):
                             continue
                     if parsed:
                         # 续行：拼回上一条 message（保留换行）
-                        prev_label, prev_body = parsed[-1]
-                        parsed[-1] = (prev_label, prev_body + "\n" + line)
+                        prev_label, prev_body, prev_legacy = parsed[-1]
+                        parsed[-1] = (prev_label, prev_body + "\n" + line, prev_legacy)
                     else:
                         # 开头就是无前缀行（异常/老数据）：当成 user 消息兜底
-                        parsed.append(("user", line))
+                        parsed.append(("user", line, _legacy))
 
             # 取最近 max_turns 条 message（按 message 截断，绝不从消息中间切开）
             recent = parsed[-max_turns:] if len(parsed) > max_turns else parsed
             if not recent:
                 return ""
 
-            # 格式化：非 user 侧整条截断（含对方真人），user 完整保留
+            # 格式化：新 raw 直接拼接；旧 raw 按写入同一规则兼容裁剪
             all_messages: List[str] = []
             if user_name:
                 from .base import render_user_name_line
@@ -3025,17 +3062,18 @@ class MemoryWriter(WritePipeline):
 
                 _name_line = render_user_name_line(
                     user_name,
-                    is_chinese("\n".join(b for _, b in recent)),
+                    is_chinese("\n".join(b for _, b, _ in recent)),
                 )
                 if _name_line:
                     all_messages.append(_name_line)
-            for label, body in recent:
+            for label, body, legacy in recent:
+                if legacy:
+                    body = trim_raw_body(
+                        label, body,
+                        assistant_max_chars=max_chars_assistant,
+                        tool_max_bytes=_tool_max_bytes,
+                    )
                 body = body.strip()
-                # 截断口径：assistant + 对方真人（非已知 role 的 label）。
-                # user 完整保留；system / tool 沿用历史行为（不截断）。
-                _truncate = label == "assistant" or label not in _KNOWN_ROLES
-                if _truncate and len(body) > max_chars_assistant:
-                    body = body[:max_chars_assistant] + "..."
                 all_messages.append(f"[{label}]: {body}")
 
             return "\n".join(all_messages)

hy_memory/utils/async_executor.py

@@ -4,6 +4,7 @@ from __future__ import annotations
 
 import asyncio
 import concurrent.futures
+import contextvars
 import functools
 from concurrent.futures import Executor
 from typing import Any, Callable, Optional
@@ -39,7 +40,9 @@ async def run_in_executor_guarded(
 
 async def to_thread_guarded(func: Callable[..., Any], *args: Any, **kwargs: Any) -> Any:
     """Guarded equivalent of ``asyncio.to_thread`` using the default pool."""
-    return await run_in_executor_guarded(None, func, *args, **kwargs)
+    context = contextvars.copy_context()
+    call = functools.partial(func, *args, **kwargs)
+    return await run_in_executor_guarded(None, context.run, call)
 
 
 def concurrent_result_guarded(

hy_memory/utils/pipeline_log_writer.py

@@ -1,6 +1,7 @@
 # -*- coding: utf-8 -*-
 """
-Pipeline Log Writer — 环节级 JSONL 日志（始终开启，与 SQLite Trace 分离）。
+Pipeline Log Writer — 环节级 JSONL 日志（默认开启，MEMORY_PIPELINE_LOG_ENABLED=false 关闭；
+与 SQLite Trace 分离）。
 
 目录: {log_dir}/{subdir}/{date}.log
 每行: request_id, namespace, user_id, agent_id, step, prompt, content, parsed,

hy_memory/utils/pipeline_observability.py

@@ -1,16 +1,20 @@
 # -*- coding: utf-8 -*-
 """
-Pipeline 可观测性：Log（文件 JSONL，始终开启）与 Trace（SQLite pipeline_logs，可关闭）。
+Pipeline 可观测性：Log（文件 JSONL，默认开启，可关闭）与 Trace（SQLite pipeline_logs，可关闭）。
 
 - Log：运维 tail/grep，每个环节即时一行，不依赖后续步骤成功与否。
 - Trace：Inspector / 调试，落库 pipeline_logs（需 MEMORY_CACHE_BACKEND=sqlite 等持久化后端）。
 
 环境变量：
+  MEMORY_PIPELINE_LOG_ENABLED    — Log 写 JSONL 文件，默认 true；false 时完全不写
   MEMORY_PIPELINE_TRACE_ENABLED  — Trace 落库，默认 true；OpenClaw C 端可设 false
   MEMORY_MEMORY_OPERATIONS_ENABLED — memory_operations 审计表落库，默认 false
   MEMORY_LOG_DIR                 — Log 目录（优先）
   MEMORY_TRACE_LOG_DIR           — Log 目录（兼容旧名）
 
+Log 与 Trace 都关闭时不包装 store_pipeline_log，而是换成立即返回 True 的
+空操作：热路径不做任何计时、脱敏或后台任务，也不写库。
+
 子目录（{log_dir}/{subdir}/{date}.log）由 PipelineLogWriter 从每条记录的
 user_id 前缀（第一个 "__" 左侧）切分；无前缀回落 "default"。该前缀是
 通用 namespace，可用于项目、部署或数据分桶。
@@ -94,6 +98,16 @@ def is_pipeline_trace_enabled() -> bool:
     )
 
 
+def is_pipeline_log_enabled() -> bool:
+    """是否把 pipeline step 写成 JSONL 文件（{log_dir}/<subdir>/<date>.log）。默认开启。"""
+    return os.getenv("MEMORY_PIPELINE_LOG_ENABLED", "true").strip().lower() in (
+        "true",
+        "1",
+        "yes",
+        "on",
+    )
+
+
 def is_memory_operations_enabled() -> bool:
     """是否将知识库变动写入 memory_operations 审计表。默认关闭（false）。"""
     return os.getenv("MEMORY_MEMORY_OPERATIONS_ENABLED", "false").lower() in (
@@ -125,7 +139,7 @@ def redact_content(value: str) -> str:
 
 def resolve_pipeline_log_dir() -> str:
     """
-    JSONL pipeline log 根目录（始终使用，有默认值）。
+    JSONL pipeline log 根目录（MEMORY_PIPELINE_LOG_ENABLED 开启时使用，有默认值）。
 
     优先级: MEMORY_LOG_DIR > MEMORY_TRACE_LOG_DIR > {MEMORY_DATA_DIR}/logs/pipeline
     """
@@ -228,6 +242,7 @@ def wrap_store_pipeline_log(
 ) -> Callable[..., Any]:
     """包装 cache.store_pipeline_log：计时同步、JSONL/DB 异步落盘，不堵主请求。
 
+    - log_writer 为 None 时不写 JSONL（MEMORY_PIPELINE_LOG_ENABLED=false）
     - step-to-step elapsed_ms / event_at_ms 在调用瞬间打点（与业务墙钟对齐）
     - event_at_ms：调用方传入非 0 值则直接复用（App 已在自己被调用的那一刻
       同步打点，比这里更接近真实事件时刻）；传 0 / 不传才用本地 time.time()
@@ -410,7 +425,8 @@ def wrap_store_pipeline_log(
         try:
             loop = asyncio.get_running_loop()
             # JSONL 用线程池，与 loop 无亲和，留在当前 loop。
-            _track_task(loop.create_task(_persist_jsonl()))
+            if log_writer is not None:
+                _track_task(loop.create_task(_persist_jsonl()))
             # DB trace 用 cache pool（可能 aiomysql，绑定 owner_loop）。若当前不在
             # owner_loop（如 search 走独立 loop），派发回 owner_loop 执行，避免跨 loop。
             if trace_enabled:
@@ -446,25 +462,26 @@ def wrap_store_pipeline_log(
                         )
         except RuntimeError:
             # 无 running loop（极少见）：退化为同步写，保证不丢日志
-            try:
-                log_writer.write_step(
-                    request_id=request_id,
-                    user_id=user_id,
-                    agent_id=agent_id,
-                    session_id=session_id,
-                    step=step,
-                    prompt=prompt,
-                    response=response,
-                    parsed=parsed,
-                    memory_ids=mem_ids,
-                    elapsed_ms=step_elapsed_ms,
-                    prompt_tokens=prompt_tokens,
-                    completion_tokens=completion_tokens,
-                    total_tokens=total_tokens,
-                    event_at_ms=event_at_ms,
-                )
-            except Exception:
-                pass
+            if log_writer is not None:
+                try:
+                    log_writer.write_step(
+                        request_id=request_id,
+                        user_id=user_id,
+                        agent_id=agent_id,
+                        session_id=session_id,
+                        step=step,
+                        prompt=prompt,
+                        response=response,
+                        parsed=parsed,
+                        memory_ids=mem_ids,
+                        elapsed_ms=step_elapsed_ms,
+                        prompt_tokens=prompt_tokens,
+                        completion_tokens=completion_tokens,
+                        total_tokens=total_tokens,
+                        event_at_ms=event_at_ms,
+                    )
+                except Exception:
+                    pass
             if trace_enabled:
                 try:
                     result = original_store(
@@ -496,6 +513,11 @@ def wrap_store_pipeline_log(
     return _store_pipeline_observability
 
 
+async def _noop_store_pipeline_log(*args: Any, **kwargs: Any) -> bool:
+    """Log 与 Trace 都关闭时的 store_pipeline_log：不计时、不写文件、不写库。"""
+    return True
+
+
 def install_pipeline_observability(
     cache: Any,
     *,
@@ -503,6 +525,9 @@ def install_pipeline_observability(
 ) -> None:
     """给 cache 安装 JSONL + Trace 包装（fire-and-forget）。
 
+    MEMORY_PIPELINE_LOG_ENABLED=false 时不写 JSONL；若 Trace 也关闭，
+    store_pipeline_log 被换成空操作（不包装、不写库）。
+
     在 cache 所属 loop（write loop / _LoopThread）上调用，本函数会捕获该 loop
     作为 owner_loop 传给包装器：search 走独立 loop 后，从 search loop 触发的 DB
     落库会被派发回这个 owner_loop，避免跨 loop 用 aiomysql 污染连接池。
@@ -514,15 +539,31 @@ def install_pipeline_observability(
     except RuntimeError:
         owner_loop = None
 
-    log_dir = resolve_pipeline_log_dir()
-    log_writer = PipelineLogWriter(log_dir)
+    log_enabled = is_pipeline_log_enabled()
     trace_enabled = is_pipeline_trace_enabled()
-    logger.info(
-        f"[{log_label}] Pipeline log (JSONL) enabled: {log_dir}/<subdir>/<date>.log "
-        f"(subdir from user_id prefix before '__', else default) ; "
-        f"pipeline trace (DB)={'on' if trace_enabled else 'off'} ; "
-        f"persist=async(fire-and-forget)"
-    )
+    if not log_enabled and not trace_enabled:
+        logger.info(
+            f"[{log_label}] Pipeline log (JSONL) disabled ; pipeline trace (DB)=off ; "
+            f"store_pipeline_log=no-op"
+        )
+        cache.store_pipeline_log = _noop_store_pipeline_log
+        return
+
+    log_writer = None
+    if log_enabled:
+        log_dir = resolve_pipeline_log_dir()
+        log_writer = PipelineLogWriter(log_dir)
+        logger.info(
+            f"[{log_label}] Pipeline log (JSONL) enabled: {log_dir}/<subdir>/<date>.log "
+            f"(subdir from user_id prefix before '__', else default) ; "
+            f"pipeline trace (DB)={'on' if trace_enabled else 'off'} ; "
+            f"persist=async(fire-and-forget)"
+        )
+    else:
+        logger.info(
+            f"[{log_label}] Pipeline log (JSONL) disabled ; "
+            f"pipeline trace (DB)=on ; persist=async(fire-and-forget)"
+        )
     cache.store_pipeline_log = wrap_store_pipeline_log(
         cache.store_pipeline_log,
         log_writer=log_writer,

hy_memory/utils/raw_trim.py

@@ -0,0 +1,91 @@
+# -*- coding: utf-8 -*-
+"""
+原始对话（raw）落库前的统一裁剪规则（规则，零 LLM）。
+
+raw 会被长期保存（MemHub 已取消 raw 的保留期限；lite 模式下 raw 是唯一存储），
+不裁剪会把工具输出、大段粘贴日志原样存下来。因此 raw 在**写入时**就裁好，
+读取方（extractor 的「最近 k 条」历史注入）直接拼接，不再二次裁剪。
+
+规则按说话人标签（ChatMessage.speaker_label() / L1_RAW 里的 ``[label]``）：
+
+- ``user``：完整保留；
+- ``system``：完整保留（与历史注入原口径一致）；
+- ``tool``：头尾保留、中段替换为 ``[...N bytes omitted by truncate...]``，
+  上限为 ``coding.tool_result_max_bytes``（env ``MEMORY_CODING_TOOL_RESULT_MAX_BYTES``，
+  默认 2048），与 coding 链路同一个函数；
+- 其余（``assistant``、对方真人名字等非已知 role 的 label）：strip 后超过
+  ``assistant_max_chars``（默认 500，即历史注入原来的 ``max_chars_assistant``）
+  时截成前 N 个字符 + ``...``。
+
+所有规则确定且幂等：对已裁剪的文本再裁一次结果不变。
+"""
+
+from typing import List, Optional
+
+from ..pipelines.base import ChatMessage
+
+# 历史注入原 `_get_recent_history(max_chars_assistant=500)` 的默认值，现为唯一来源。
+DEFAULT_RAW_ASSISTANT_MAX_CHARS = 500
+
+# L1_RAW 节点 custom 上的标记：内容已按本模块规则在写入时裁剪。
+# 历史拼接据此区分「新 raw 直接拼接」与「本规则之前写入的旧 raw 需兼容裁剪」。
+RAW_TRIMMED_KEY = "raw_trimmed"
+
+# 不裁剪的标签（tool 单独处理）
+_KEEP_LABELS = frozenset({"user", "system"})
+
+
+def trim_raw_body(
+    label: str,
+    body: str,
+    *,
+    assistant_max_chars: int = DEFAULT_RAW_ASSISTANT_MAX_CHARS,
+    tool_max_bytes: Optional[int] = None,
+) -> str:
+    """按说话人标签裁剪单条消息正文。幂等。"""
+    if not isinstance(body, str):
+        body = "" if body is None else str(body)
+    if label in _KEEP_LABELS:
+        return body
+    if label == "tool":
+        from ..coding.preproc import (
+            DEFAULT_MAX_TOOL_RESULT_BYTES,
+            truncate_tool_result_text,
+        )
+
+        return truncate_tool_result_text(
+            body, max_bytes=int(tool_max_bytes or DEFAULT_MAX_TOOL_RESULT_BYTES),
+        )
+    stripped = body.strip()
+    if len(stripped) > assistant_max_chars:
+        return stripped[:assistant_max_chars] + "..."
+    return body
+
+
+def trim_raw_messages(
+    messages: List[ChatMessage],
+    *,
+    assistant_max_chars: int = DEFAULT_RAW_ASSISTANT_MAX_CHARS,
+    tool_max_bytes: Optional[int] = None,
+) -> List[ChatMessage]:
+    """返回裁剪后的新列表；不修改入参（messages 仍要原样给 extractor）。"""
+    out: List[ChatMessage] = []
+    for m in messages:
+        new_content = trim_raw_body(
+            m.speaker_label(),
+            m.content,
+            assistant_max_chars=assistant_max_chars,
+            tool_max_bytes=tool_max_bytes,
+        )
+        if new_content == m.content:
+            out.append(m)
+            continue
+        out.append(ChatMessage(
+            role=m.role,
+            content=new_content,
+            name=m.name,
+            tool_calls=list(m.tool_calls),
+            tool_call_id=m.tool_call_id,
+            tool_name=m.tool_name,
+        ))
+    return out

hy_memory/utils/token_count.py

@@ -36,7 +36,8 @@ def count_tokens(text: str) -> int:
     enc = _get_encoder()
     if enc:
         try:
-            return len(enc.encode(text))
+            # Quoted control-token spellings are ordinary input text.
+            return len(enc.encode(text, disallowed_special=()))
         except Exception as e:
             logger.debug(f"[token_count] encode failed, char fallback: {e}")
     return max(1, len(text) // 4)

memhub/MEMORY.md

@@ -13,8 +13,8 @@
    接口先补 SDK。Runtime writers 停止后的 App 启动迁移可直接操作底层存储，
    但不得把迁移专用方法加入公共 Client。
 3. MemHub **不存 memory 实体**；**可以存 `memory_id`** 做组织/回执。原文进入 **SDK raw**。
-4. SDK raw 只保留滚动三天；一条 raw 对应一个完整 turn，只含真人 user 与
-   assistant，tool 不持久化。提取后的高层 memory 不受该窗口影响。
+4. SDK raw 全部保留，任何模式（lite / pro / ultra）都不删除；一条 raw 对应一个
+   完整 turn，只含真人 user 与 assistant，tool 不持久化。
 5. `source_raw_ids` 是 SDK 内部溯源关系，不是产品 Evidence 合同；记忆详情不提供
    “原始依据”或按 raw ID 展开原文的入口。
 

memhub/PRIVACY.md

@@ -29,10 +29,12 @@ prompt/response、密钥、访问令牌或完整本地路径。结构化日志
 文件。详细合同见
 [`docs/contracts/diagnostic-logging.md`](docs/contracts/diagnostic-logging.md)。
 
-用户显式把诊断级别切换为 `debug` 时，Core 会启用 HY pipeline trace，并将
-`MEMORY_LOG_LLM_CONTENT` 设为 `true`。该模式会把完整模型 prompt/response 写入本地调试
-日志，其中可能包含获授权的对话正文和模型输出；它只应用于有意识的故障排查，不应长期
-开启，也不应在未经检查和脱敏时分享相关日志。切回 `info` 会停止后续正文记录，但不会
+HY pipeline 日志与 trace 在任何级别下都关闭。HY SDK 的文本日志只写入
+`hy-memory/governed-v2/logs/hy_memory.log`，其中的 WARNING 及以上同时写入 `core.log`；
+SDK 不在 INFO 及以上级别记录对话或记忆正文。用户显式把诊断级别切换为 `debug` 时，Core 将
+`MEMORY_LOG_LLM_CONTENT` 设为 `true`。该模式会把完整模型 prompt/response 写入
+`hy_memory.log` 的 DEBUG 行，其中可能包含获授权的对话正文和模型输出；它只应用于有意识的
+故障排查，不应长期开启，也不应在未经检查和脱敏时分享相关日志。切回 `info` 会停止后续正文记录，但不会
 自动删除此前已写入的调试文件。
 
 具体行为以当前源码、网络请求实现和发布说明为准。发现实现与本说明不一致时，请按

memhub/README.md

@@ -18,13 +18,13 @@ If you run into a problem, please open an Issue with the version, platform,repro
 
 ## 发行与平台合同
 
-源码当前固定使用公开 PyPI 正式版 `hy-memory==2.0.6`。MemHub 源码版本为 `1.0.5`
-（build 14），GitHub 仓库为
+源码当前固定使用公开 PyPI 正式版 `hy-memory==2.0.9`。MemHub 源码版本为 `1.0.5`
+（build 36），GitHub 仓库为
 [`Tencent-Hunyuan/Hy-Memory`](https://github.com/Tencent-Hunyuan/Hy-Memory)，npm 包名为
 `memhub-cli`。当前代码架构见
 [`docs/ARCHITECTURE.md`](docs/ARCHITECTURE.md)。
 
-`v1.0.5` 的发行目标是 Linux x64/arm64 Runtime 与 macOS arm64/x64 App 和 Runtime。
+`v1.0.5` 的发行目标是 Linux x64/arm64 Runtime 与 macOS arm64 App 和 Runtime。
 Linux 制品由 GitHub Actions 在原生架构 runner 上构建并完成生命周期验收；macOS
 制品由发行流水线完成 Developer ID 签名、公证和验收后手动上传。Windows 不支持。
 只有同时出现在 GitHub Release 与 `release-manifest.json` 中、且通过公开校验的制品才属于
@@ -118,7 +118,7 @@ MemHub 的 SQLite 数据默认位于 `~/.one-memory`，不在 npm 包或 Runtime
 
 ## Hy Memory
 
-SDK 依赖固定为公开 PyPI 正式版 `hy-memory==2.0.6`。SDK 配置、权限边界及 search
+SDK 依赖固定为公开 PyPI 正式版 `hy-memory==2.0.9`。SDK 配置、权限边界及 search
 返回格式见 [`MEMORY.md`](MEMORY.md)。HY search 返回的 `content` 已包含 SDK 统一生成的
 逻辑时间和演化上下文；调用方不应再次转换时间戳或重复拼接 `When`。
 

memhub/SBOM.spdx.json

@@ -2352,7 +2352,7 @@
     {
       "SPDXID": "SPDXRef-Package-hy-memory",
       "name": "hy-memory",
-      "versionInfo": "2.0.6",
+      "versionInfo": "2.0.9",
       "downloadLocation": "NOASSERTION",
       "filesAnalyzed": false,
       "licenseConcluded": "Apache-2.0",

memhub/THIRD_PARTY_NOTICES.md

@@ -9,7 +9,7 @@ Per-package license texts are not reproduced here yet; the license
 column is filled in as part of release review, and the SBOM records
 `NOASSERTION` until then rather than guessing.
 
-## hy-memory 2.0.6
+## hy-memory 2.0.9
 
 License: Apache-2.0
 

memhub/apps/cli/src/cli-contract-fixes.test.ts

@@ -315,6 +315,8 @@ describe("catalog exit codes", () => {
       "models",
       "memories",
       "profile",
+      "settings",
+      "evolve",
       "graph-width",
       "host",
       "recall",

memhub/apps/cli/src/cli-logs.test.ts

@@ -162,10 +162,54 @@ it("never emits a secret from a malformed manually written event", async () => {
     command: "logs",
     state: "empty",
     events: 0,
-    invalid_lines: 1,
+    invalid_lines: 0,
+    unrecognized_lines: 1,
   });
 });
 
+// After an upgrade the files still hold an older build's records. They are
+// whole JSON lines the reader leaves out, not damage, so the text output must
+// not warn about them; only a line that is not JSON is called damaged.
+it("warns only about lines that are not JSON, not about older records", async () => {
+  const dataDir = await fixture();
+  const older = {
+    schema_version: 0,
+    service: "desktop",
+    component: "desktop.lifecycle",
+    event: "desktop.started",
+  };
+  await writeFile(
+    path.join(dataDir, "logs", "diagnostic-desktop.jsonl"),
+    `${JSON.stringify(older)}\n${JSON.stringify({ ...older, schema_version: 1 })}\n`,
+    { mode: 0o600 },
+  );
+  let stdout = "";
+  let stderr = "";
+  vi.spyOn(process.stdout, "write").mockImplementation((value) => {
+    stdout += String(value);
+    return true;
+  });
+  vi.spyOn(process.stderr, "write").mockImplementation((value) => {
+    stderr += String(value);
+    return true;
+  });
+
+  expect(await run(["logs"], { ONE_MEMORY_DATA_DIR: dataDir })).toBe(0);
+  expect(stdout).toContain("core.start_failed");
+  expect(stdout).not.toContain("损坏");
+  expect(stderr).toBe("");
+
+  await writeFile(
+    path.join(dataDir, "logs", "diagnostic-cli.jsonl"),
+    '{"schema_version": 1, "event": \n',
+    { mode: 0o600 },
+  );
+  stdout = "";
+  expect(await run(["logs"], { ONE_MEMORY_DATA_DIR: dataDir })).toBe(0);
+  expect(stdout).toContain("跳过 1 行损坏的诊断记录。");
+  expect(stderr).toBe("memhub: 跳过 1 行损坏的诊断记录。\n");
+});
+
 it("rejects invalid filters without touching the log directory", async () => {
   const dataDir = path.join(
     await mkdtemp(path.join(os.tmpdir(), "om-cli-no-log-")),

memhub/apps/cli/src/cli.ts

@@ -188,10 +188,31 @@ import {
   removeManagedCliLink,
   rollbackManagedCliLink,
 } from "./managed-cli-link.js";
+import {
+  type FeaturesDocument,
+  featuresExitCode,
+  parseFeaturesCommand,
+  renderFeaturesText,
+  runFeaturesCommand,
+} from "./memory-features-flow.js";
+import {
+  type EvolveDocument,
+  evolveExitCode,
+  type ModeDocument,
+  modeExitCode,
+  parseEvolveCommand,
+  parseSettingsCommand,
+  renderEvolveText,
+  renderModeText,
+  runEvolveCommand,
+  runModeCommand,
+} from "./memory-mode-flow.js";
 import {
   addMemory,
+  DEFAULT_USER_ID,
   isMemoryId,
   isRequestId,
+  isUserId,
   MEMORY_KINDS,
   MEMORY_SCENES,
   MEMORY_VARIANTS,
@@ -209,6 +230,7 @@ import {
   renderMemoryAddText,
   renderMemoryDetailText,
   renderMemoryPageText,
+  userIdParameters,
   validMemoryText,
 } from "./memory-report.js";
 import {
@@ -322,6 +344,8 @@ export interface RunControls {
   watchSignal?: AbortSignal;
   watchWait?: (milliseconds: number, signal: AbortSignal) => Promise<void>;
   watchNow?: () => string;
+  /** `memhub evolve --wait` polls on this; tests inject an instant one. */
+  sleep?: (milliseconds: number) => Promise<void>;
   readStdin?: () => Promise<Buffer | null>;
   stdinIsTTY?: boolean;
   /** Injectable onboarding prompter so the wizard can be driven in tests
@@ -538,6 +562,23 @@ async function runCommand(
       environment = { ...environment, ONE_MEMORY_DATA_DIR: lifted.dataDir };
     }
   }
+  // `--user-id <id>` is global to the memory commands: it names whose
+  // memories are read or written. It is lifted before dispatch so each
+  // command's own parser keeps seeing exactly the arguments it always did.
+  // `sources connect` gives the flag its own meaning and parses it itself.
+  let userId: string | undefined;
+  if (USER_ID_COMMANDS.has(arguments_[0] ?? "")) {
+    const lifted = liftUserIdOption(arguments_);
+    if (lifted === null) {
+      return reject(
+        findCommand(arguments_[0] ?? ""),
+        asJson,
+        "--user-id 无效：只能是 1–64 位字母、数字、点、下划线或连字符。",
+      );
+    }
+    arguments_ = lifted.arguments_;
+    userId = lifted.userId;
+  }
   environment = await environmentWithManagedDataDir(environment);
   const command = arguments_[0];
   if (command === undefined) return reject(undefined, asJson, "缺少命令。");
@@ -776,9 +817,11 @@ async function runCommand(
     const attachment = await attachToRuntimeOrNull(environment);
     if (attachment === null) return runtimeUnavailable(descriptor, asJson);
     if (parsed.kind === "status") {
-      const receipt = await readMemoryWrite(attachment, parsed.id).catch(
-        () => undefined,
-      );
+      const receipt = await readMemoryWrite(
+        attachment,
+        parsed.id,
+        userId,
+      ).catch(() => undefined);
       if (receipt === undefined) {
         return emitFailure(asJson, "write status", "memory_write_read_failed");
       }
@@ -812,6 +855,7 @@ async function runCommand(
         ? {}
         : { memory_at: parsed.memory_at }),
       ...(parsed.priority === undefined ? {} : { priority: parsed.priority }),
+      ...(userId === undefined ? {} : { user_id: userId }),
     });
     if (outcome.state !== "accepted") {
       emit(
@@ -957,9 +1001,12 @@ async function runCommand(
                 state: "empty",
                 events: 0,
                 invalid_lines: report.invalid_lines,
+                unrecognized_lines: report.unrecognized_lines,
               })}\n`
           : renderLogText(report.events, report.invalid_lines),
       );
+      // Only lines that are not JSON are damage. Records an older build wrote
+      // are left out without a warning: after an upgrade they are expected.
       if (report.invalid_lines) {
         process.stderr.write(
           `${PRODUCT}: 跳过 ${report.invalid_lines} 行损坏的诊断记录。\n`,
@@ -1582,6 +1629,7 @@ async function runCommand(
           limit: parsed.limit,
           character_budget: parsed.characterBudget,
           timeout_ms: parsed.timeoutMs,
+          ...userIdParameters(userId),
         })
       : ({
           schema_version: 2,
@@ -1780,6 +1828,16 @@ async function runCommand(
   }
   if (named === "sources" && arguments_[1] !== "list") {
     const parsed = parseSourcesCommand(arguments_.slice(1));
+    if (
+      !parsed &&
+      arguments_.some((item) => item === "--user-id" || item === "--root")
+    ) {
+      return reject(
+        descriptor,
+        asJson,
+        "--user-id 和 --root 只能用于接入单个 hermes 来源（Hermes profile），用户 ID 只能是 1–64 位字母、数字、点、下划线或连字符。",
+      );
+    }
     if (!parsed) return usage();
     if (parsed.kind === "receipt") {
       const connected = await connectRuntime(environment);
@@ -1981,12 +2039,17 @@ async function runCommand(
     return sourceProgressExitCode(document);
   }
   if (named === "profile") {
+    const profileFlags = arguments_.slice(1);
     if (
-      arguments_.filter((item) => item === "--json").length > 1 ||
-      arguments_.slice(1).some((value) => value !== "--json")
+      profileFlags.filter((item) => item === "--json").length > 1 ||
+      profileFlags.filter((item) => item === "--versions").length > 1 ||
+      profileFlags.some((value) => value !== "--json" && value !== "--versions")
     ) {
       return usage();
     }
+    if (profileFlags.includes("--versions")) {
+      return await profileVersionsCommand(environment, { asJson, userId });
+    }
     const connected = await connectRuntime(environment, {
       readerTimeoutMs: 15_000,
     });
@@ -2004,7 +2067,11 @@ async function runCommand(
       );
       return connected.exitCode;
     }
-    const outcome = await connected.reader<unknown>("/api/v2/user-profile");
+    const profileUser = userIdParameters(userId);
+    const outcome =
+      profileUser.user_id === undefined
+        ? await connected.reader<unknown>("/api/v2/user-profile")
+        : await connected.reader<unknown>("/api/v2/user-profile", profileUser);
     const value =
       outcome.state === "ok"
         ? (outcome.data as {
@@ -2109,7 +2176,7 @@ async function runCommand(
     }
     const { reader, operations } = connected;
     if (parsed.kind === "list") {
-      const document = await readMemoryPage(reader, parsed.query);
+      const document = await readMemoryPage(reader, parsed.query, userId);
       emit(
         asJson
           ? `${JSON.stringify(document)}\n`
@@ -2122,6 +2189,7 @@ async function runCommand(
         reader,
         parsed.memoryId,
         parsed.namespace,
+        userId,
       );
       emit(
         asJson
@@ -2133,8 +2201,8 @@ async function runCommand(
     if (parsed.kind === "add" || parsed.kind === "add-status") {
       const document =
         parsed.kind === "add"
-          ? await addMemory(operations, text ?? "", parsed.requestId)
-          : await readMemoryAddStatus(reader, parsed.requestId);
+          ? await addMemory(operations, text ?? "", parsed.requestId, userId)
+          : await readMemoryAddStatus(reader, parsed.requestId, userId);
       emit(
         asJson
           ? `${JSON.stringify(document)}\n`
@@ -2149,9 +2217,13 @@ async function runCommand(
             content: text,
             ...namespaceParameters(parsed.namespace),
             ...(parsed.task === undefined ? {} : { task: parsed.task }),
+            ...userIdParameters(userId),
           })
         : await operations.delete<unknown>(
-            withQuery(memoryPath, namespaceParameters(parsed.namespace)),
+            withQuery(memoryPath, {
+              ...namespaceParameters(parsed.namespace),
+              ...userIdParameters(userId),
+            }),
           );
     const document = {
       schema_version: 2,
@@ -2177,6 +2249,89 @@ async function runCommand(
     );
     return outcome.state === "committed" ? EXIT_OK : EXIT_FAILED;
   }
+  if (
+    named === "settings" &&
+    arguments_.slice(1).find((value) => value !== "--json") === "features"
+  ) {
+    const parsed = parseFeaturesCommand(
+      arguments_
+        .slice(1)
+        .filter((value, index, all) => index !== all.indexOf("features")),
+    );
+    if (!parsed) return usage();
+    const connected = await connectRuntime(environment);
+    if (!connected.ok) {
+      const document: FeaturesDocument = {
+        schema_version: 1,
+        product: PRODUCT,
+        command: "settings features",
+        state: "unavailable",
+        code: connected.code,
+      };
+      emit(
+        asJson ? `${JSON.stringify(document)}\n` : renderFeaturesText(document),
+      );
+      return connected.exitCode;
+    }
+    const document = await runFeaturesCommand(
+      parsed,
+      connected.reader,
+      connected.operations,
+    );
+    emit(
+      asJson ? `${JSON.stringify(document)}\n` : renderFeaturesText(document),
+    );
+    return featuresExitCode(document);
+  }
+  if (named === "settings") {
+    const parsed = parseSettingsCommand(arguments_.slice(1));
+    if (!parsed) return usage();
+    const connected = await connectRuntime(environment);
+    if (!connected.ok) {
+      const document: ModeDocument = {
+        schema_version: 1,
+        product: PRODUCT,
+        command: "settings mode",
+        state: "unavailable",
+        code: connected.code,
+      };
+      emit(asJson ? `${JSON.stringify(document)}\n` : renderModeText(document));
+      return connected.exitCode;
+    }
+    const document = await runModeCommand(
+      parsed,
+      connected.reader,
+      connected.operations,
+    );
+    emit(asJson ? `${JSON.stringify(document)}\n` : renderModeText(document));
+    return modeExitCode(document);
+  }
+  if (named === "evolve") {
+    const parsed = parseEvolveCommand(arguments_.slice(1));
+    if (!parsed) return usage();
+    const connected = await connectRuntime(environment);
+    if (!connected.ok) {
+      const document: EvolveDocument = {
+        schema_version: 1,
+        product: PRODUCT,
+        command: "evolve",
+        state: "unavailable",
+        code: connected.code,
+      };
+      emit(
+        asJson ? `${JSON.stringify(document)}\n` : renderEvolveText(document),
+      );
+      return connected.exitCode;
+    }
+    const document = await runEvolveCommand(parsed, {
+      userId: userIdParameters(userId).user_id,
+      reader: connected.reader,
+      operations: connected.operations,
+      sleep: controls.sleep,
+    });
+    emit(asJson ? `${JSON.stringify(document)}\n` : renderEvolveText(document));
+    return evolveExitCode(document);
+  }
   if (named === "graph-width") {
     const parsed = parseGraphWidthCommand(arguments_.slice(1));
     if (!parsed) return usage();
@@ -3675,13 +3830,21 @@ async function connectSources(
       asJson,
     );
   }
-  if (parsed.importHistory) {
+  // Another user's Hermes profile is granted and imported in one connect
+  // request: only that request can carry a user and a root.
+  if (
+    parsed.importHistory ||
+    parsed.userId !== undefined ||
+    parsed.root !== undefined
+  ) {
     const document = await connectSourcesWithImport(reader, operations, {
       sourceIds: parsed.sourceIds,
       confirmOverrides: parsed.confirmOverrides,
       ...(parsed.requestId === undefined
         ? {}
         : { requestId: parsed.requestId }),
+      ...(parsed.userId === undefined ? {} : { userId: parsed.userId }),
+      ...(parsed.root === undefined ? {} : { root: parsed.root }),
     });
     emitFlowDocument(document, asJson);
     return flowExitCode(document);
@@ -3834,6 +3997,108 @@ async function emitModelReceipt(
 
 /** Attach to a verified Runtime, telling a stopped one (exit 3) from one
  * that is up but unusable (exit 1). */
+/**
+ * `memhub profile --versions`: the long-term profile's L0 version chains,
+ * latest first, so a value that replaced a better one can be found by id.
+ */
+async function profileVersionsCommand(
+  environment: NodeJS.ProcessEnv,
+  options: { asJson: boolean; userId?: string },
+): Promise<number> {
+  const fail = (state: string, code: string, exitCode: number): number => {
+    emit(
+      options.asJson
+        ? `${JSON.stringify({ schema_version: 2, product: PRODUCT, state, code })}\n`
+        : `画像版本读取失败：${code}\n`,
+    );
+    return exitCode;
+  };
+  const connected = await connectRuntime(environment, {
+    readerTimeoutMs: 15_000,
+  });
+  if (!connected.ok) {
+    return fail("unavailable", connected.code, connected.exitCode);
+  }
+  const outcome = await connected.reader<unknown>(
+    "/api/v2/user-profile/versions",
+    userIdParameters(options.userId),
+  );
+  const chains = outcome.state === "ok" ? profileChains(outcome.data) : null;
+  if (outcome.state !== "ok" || chains === null) {
+    return fail(
+      outcome.state === "ok" ? "invalid" : outcome.state,
+      outcome.state === "ok" ? "response_invalid" : outcome.code,
+      EXIT_FAILED,
+    );
+  }
+  if (options.asJson) {
+    emit(
+      `${JSON.stringify({
+        schema_version: 2,
+        product: PRODUCT,
+        state: "ok",
+        ...(outcome.data as object),
+      })}\n`,
+    );
+    return EXIT_OK;
+  }
+  if (chains.length === 0) {
+    emit("尚未形成用户画像版本。\n");
+    return EXIT_OK;
+  }
+  const lines: string[] = [];
+  for (const chain of chains) {
+    lines.push(`画像 ${chain.memory_id}（${chain.status}）`);
+    for (const version of chain.versions) {
+      const when =
+        typeof version.gmt_created === "number"
+          ? new Date(version.gmt_created * 1000).toISOString()
+          : "时间未知";
+      lines.push(
+        `  ${version.current ? "*" : " "} ${version.memory_id}  ${when}  ${version.content}`,
+      );
+    }
+  }
+  lines.push(
+    "* 表示当前生效的版本；删除错误版本：memhub memories delete <memory-id> --yes",
+  );
+  emit(`${lines.join("\n")}\n`);
+  return EXIT_OK;
+}
+
+type ProfileChainView = {
+  memory_id: string;
+  status: string;
+  versions: {
+    memory_id: string;
+    content: string;
+    gmt_created?: number | null;
+    current: boolean;
+  }[];
+};
+
+function profileChains(value: unknown): ProfileChainView[] | null {
+  if (!value || typeof value !== "object") return null;
+  const chains = (value as { chains?: unknown }).chains;
+  if (!Array.isArray(chains)) return null;
+  const valid = chains.every(
+    (chain) =>
+      chain &&
+      typeof chain === "object" &&
+      typeof (chain as ProfileChainView).memory_id === "string" &&
+      typeof (chain as ProfileChainView).status === "string" &&
+      Array.isArray((chain as ProfileChainView).versions) &&
+      (chain as ProfileChainView).versions.every(
+        (version) =>
+          version &&
+          typeof version.memory_id === "string" &&
+          typeof version.content === "string" &&
+          typeof version.current === "boolean",
+      ),
+  );
+  return valid ? (chains as ProfileChainView[]) : null;
+}
+
 async function connectRuntime(
   environment: NodeJS.ProcessEnv,
   options?: { readerTimeoutMs?: number },
@@ -4101,6 +4366,8 @@ const SDK_MIGRATION_FAILURES: Record<string, string> = {
     "核对发现新版本还读不到你的全部记忆，没有切换数据源。",
   sdk_migration_model_profile_required:
     "迁移需要先有可用的模型配置，用来核对新版本能否读回记忆。请先用 memhub models setup 完成配置，再运行 memhub restart 重试。",
+  sdk_migration_vector_profile_missing:
+    "记忆的向量来自一个已删除的模型配置，当前配置读不回它们，所以没有迁移。请重新添加并启用那个模型配置，再运行 memhub restart 重试。",
 };
 
 const MIGRATION_ISSUES_URL =
@@ -5131,14 +5398,23 @@ type SourcesCommand =
       importHistory: boolean;
       confirmOverrides: boolean;
       requestId?: string;
+      /** `--user-id`: whose grant (Hermes only). Absent for `local`. */
+      userId?: string;
+      /** `--root`: an absolute root other than the candidate (Hermes only). */
+      root?: string;
     }
   | { kind: "all"; importHistory: boolean; confirmOverrides: boolean }
   | { kind: "receipt"; requestId: string };
 
 const SOURCE_ID_SHAPE = /^[a-z0-9][a-z0-9-]{0,127}$/;
 
+/** Only Hermes keeps one history per profile, so only Hermes may be
+ * connected for another user or from another root. */
+const MULTI_USER_SOURCES = new Set(["hermes"]);
+
 /** `sources connect <id> [<id> ...] | --all-detected [--import]
- * [--confirm-overrides] [--request-id <uuid>]` and `sources receipt <id>`.
+ * [--confirm-overrides] [--request-id <uuid>]
+ * [--user-id <id>] [--root <path>]` and `sources receipt <id>`.
  *
  * Sources may be named positionally or with repeated `--source` (the old
  * `init --source` spelling is rewritten to this command). */
@@ -5154,6 +5430,8 @@ function parseSourcesCommand(rest: readonly string[]): SourcesCommand | null {
   const sourceIds: string[] = [];
   const switches = new Set<string>();
   let requestId: string | undefined;
+  let userId: string | undefined;
+  let root: string | undefined;
   for (let index = 1; index < tokens.length; index += 1) {
     const token = tokens[index];
     if (["--import", "--confirm-overrides", "--all-detected"].includes(token)) {
@@ -5161,6 +5439,21 @@ function parseSourcesCommand(rest: readonly string[]): SourcesCommand | null {
       switches.add(token);
       continue;
     }
+    if (token === "--user-id" || token === "--root") {
+      const value = tokens[index + 1];
+      if (value === undefined || value.startsWith("--")) return null;
+      if (token === "--user-id") {
+        if (userId !== undefined || !isUserId(value)) return null;
+        userId = value;
+      } else {
+        if (root !== undefined) return null;
+        const expanded = expandRootPath(value);
+        if (expanded === null) return null;
+        root = expanded;
+      }
+      index += 1;
+      continue;
+    }
     if (token === "--source" || token === "--request-id") {
       const value = tokens[index + 1];
       if (value === undefined || value.startsWith("--")) return null;
@@ -5184,10 +5477,23 @@ function parseSourcesCommand(rest: readonly string[]): SourcesCommand | null {
   const confirmOverrides = switches.has("--confirm-overrides");
   if (
     requestId !== undefined &&
-    (!OPERATION_ID.test(requestId) || !importHistory)
+    (!OPERATION_ID.test(requestId) ||
+      !(importHistory || userId !== undefined || root !== undefined))
   ) {
     return null;
   }
+  // `--user-id local` is the default and changes nothing.
+  if (userId === DEFAULT_USER_ID) userId = undefined;
+  if (userId !== undefined || root !== undefined) {
+    // One Hermes profile per connect: the user and root name one grant.
+    if (
+      switches.has("--all-detected") ||
+      sourceIds.length !== 1 ||
+      !MULTI_USER_SOURCES.has(sourceIds[0])
+    ) {
+      return null;
+    }
+  }
   if (switches.has("--all-detected")) {
     // One names exact sources, the other means "whatever is detected".
     if (sourceIds.length > 0 || requestId !== undefined) return null;
@@ -5200,9 +5506,24 @@ function parseSourcesCommand(rest: readonly string[]): SourcesCommand | null {
     importHistory,
     confirmOverrides,
     ...(requestId === undefined ? {} : { requestId }),
+    ...(userId === undefined ? {} : { userId }),
+    ...(root === undefined ? {} : { root }),
   };
 }
 
+/** `--root` as an absolute path: a leading `~` is the home directory, a
+ * relative path is taken from the working directory. Core expands nothing. */
+function expandRootPath(value: string): string | null {
+  if (value.length === 0 || /[\0\r\n]/.test(value)) return null;
+  const expanded =
+    value === "~"
+      ? os.homedir()
+      : value.startsWith("~/")
+        ? path.join(os.homedir(), value.slice(2))
+        : value;
+  return path.resolve(expanded);
+}
+
 function parseGraphWidthCommand(
   rest: readonly string[],
 ): { decision: GraphWidthDecision; confirmed: boolean } | null {
@@ -6596,6 +6917,36 @@ const OWN_DATA_DIR_COMMANDS = new Set([
 
 /** Remove one `--data-dir <absolute path>` from argv; null when it is
  * repeated, missing its value, or not an absolute path. */
+/** Memory commands that honour the global `--user-id`. */
+const USER_ID_COMMANDS = new Set([
+  "memories",
+  "recall",
+  "profile",
+  "write",
+  "evolve",
+]);
+
+/**
+ * Take `--user-id <id>` out of the arguments. The default user (`local`) is
+ * returned as undefined, so a default request names no user at all and stays
+ * byte-identical to one sent before multi-user support.
+ */
+function liftUserIdOption(
+  arguments_: readonly string[],
+): { arguments_: string[]; userId?: string } | null {
+  const index = arguments_.indexOf("--user-id");
+  if (index === -1) return { arguments_: [...arguments_] };
+  if (arguments_.indexOf("--user-id", index + 1) !== -1) return null;
+  const value = arguments_[index + 1];
+  if (value === undefined || value.startsWith("--") || !isUserId(value)) {
+    return null;
+  }
+  return {
+    arguments_: [...arguments_.slice(0, index), ...arguments_.slice(index + 2)],
+    ...(value === DEFAULT_USER_ID ? {} : { userId: value }),
+  };
+}
+
 function liftDataDirOption(
   arguments_: readonly string[],
 ): { arguments_: string[]; dataDir?: string } | null {

memhub/apps/cli/src/command-catalog.ts

@@ -68,6 +68,12 @@ const JSON_OPTION: CommandOption = {
   flag: "--json",
   summary: "以 JSON 输出到 stdout，供脚本与 Agent 消费。",
 };
+/** Global to the memory commands: whose memories are read or written. */
+const USER_ID_OPTION: CommandOption = {
+  flag: "--user-id <id>",
+  summary:
+    "读写哪个用户的记忆，默认 local（App 用的就是 local）。1–64 位字母、数字、点、下划线或连字符；其他用户的记忆来自用 sources connect hermes --user-id 接入的 Hermes profile。",
+};
 const WATCH_OPTION: CommandOption = {
   flag: "--watch",
   summary: "串行持续刷新；Ctrl-C 只停止观察，不停止 Runtime。",
@@ -252,15 +258,25 @@ export const COMMANDS: readonly CommandDescriptor[] = [
   {
     name: "sources",
     summary:
-      "列出来源，直接接入来源，或按 request id 查接入回执。connect 不加 --import 只授权（处理保持冻结：不导入历史、不扫描新会话）；加 --import 同时导入历史并开启持续同步，返回 request id。目录来自环境变量覆盖的来源必须加 --confirm-overrides。",
+      "列出来源，直接接入来源，或按 request id 查接入回执。connect 不加 --import 只授权（处理保持冻结：不导入历史、不扫描新会话）；加 --import 同时导入历史并开启持续同步，返回 request id。目录来自环境变量覆盖的来源必须加 --confirm-overrides。Hermes 的每个 profile 是一个独立用户：用 --user-id 和 --root 接入，授权并导入历史（会自动排除 ~/.hermes/profiles/**，各用户互不读取）。",
     usage:
-      "memhub sources list [--json]\n            memhub sources connect <id> [<id> ...] [--import] [--confirm-overrides] [--request-id <uuid>] [--json]\n            memhub sources connect --all-detected [--import] [--confirm-overrides] [--json]\n            memhub sources receipt <request-id> [--json]",
+      "memhub sources list [--json]\n            memhub sources connect <id> [<id> ...] [--import] [--confirm-overrides] [--request-id <uuid>] [--json]\n            memhub sources connect hermes --user-id <id> --root <path> [--import] [--request-id <uuid>] [--json]\n            memhub sources connect --all-detected [--import] [--confirm-overrides] [--json]\n            memhub sources receipt <request-id> [--json]",
     options: [
       {
         flag: "--import",
         summary:
           "接入后立即导入历史会话并开启持续同步；不加时只授权，处理保持冻结。",
       },
+      {
+        flag: "--user-id <id>",
+        summary:
+          "只用于 hermes：这条接入属于哪个用户（默认 local）。和 --root 一起接入一个 Hermes profile，例如 --user-id wife --root ~/.hermes/profiles/wife；其他来源只能是 local。",
+      },
+      {
+        flag: "--root <path>",
+        summary:
+          "只用于 hermes：读取的根目录（~ 会展开为主目录）。默认是 Hermes 自己的目录 ~/.hermes。",
+      },
       {
         flag: "--all-detected",
         summary:
@@ -373,7 +389,7 @@ export const COMMANDS: readonly CommandDescriptor[] = [
     summary:
       "列出、搜索、查看、添加、修改或物理删除记忆（对话、工作、编码三个场景）。add 和 App 的「添加记忆」是同一个入口；delete 不可逆，需 --yes。",
     usage:
-      "memhub memories list [--scene <all|chat|working|coding>] [--source <id>] [--query <text>] [--status <state>] [--offset <n>] [--limit <n>] [--json]\n            memhub memories show <memory-id> [--kind <kind>] [--variant <variant>] [--json]\n            memhub memories add <text> [--request-id <uuid>] [--json]\n            memhub memories add --content-file <path> [--request-id <uuid>] [--json]\n            memhub memories add-status <request-id> [--json]\n            memhub memories update <memory-id> --content-file <path> [--kind <kind>] [--variant <variant>] [--task <text>] [--json]\n            memhub memories delete <memory-id> [--kind <kind>] [--variant <variant>] [--yes] [--json]",
+      "memhub memories list [--scene <all|chat|working|coding>] [--source <id>] [--query <text>] [--status <state>] [--offset <n>] [--limit <n>] [--user-id <id>] [--json]\n            memhub memories show <memory-id> [--kind <kind>] [--variant <variant>] [--user-id <id>] [--json]\n            memhub memories add <text> [--request-id <uuid>] [--user-id <id>] [--json]\n            memhub memories add --content-file <path> [--request-id <uuid>] [--user-id <id>] [--json]\n            memhub memories add-status <request-id> [--user-id <id>] [--json]\n            memhub memories update <memory-id> --content-file <path> [--kind <kind>] [--variant <variant>] [--task <text>] [--user-id <id>] [--json]\n            memhub memories delete <memory-id> [--kind <kind>] [--variant <variant>] [--yes] [--user-id <id>] [--json]",
     options: [
       {
         flag: "--scene <all|chat|working|coding>",
@@ -411,15 +427,75 @@ export const COMMANDS: readonly CommandDescriptor[] = [
           "add 的幂等键；用同一个 id 重试只会读回第一次的结果，不会重复添加。不给时自动生成并输出。",
       },
       DESTRUCTIVE_CONFIRM_OPTION,
+      USER_ID_OPTION,
       JSON_OPTION,
     ],
     exit_codes: RUNTIME_EXITS,
   },
   {
     name: "profile",
-    summary: "读取 SDK 组装后的用户长期画像（L0 + 已启用的长期画像补充）。",
-    usage: "memhub profile [--json]",
-    options: [JSON_OPTION],
+    summary:
+      "读取 SDK 组装后的用户长期画像（L0 + 已启用的长期画像补充）；--versions 列出 L0 画像的版本链（最新在前，* 为当前版本）。",
+    usage: "memhub profile [--versions] [--user-id <id>] [--json]",
+    options: [
+      {
+        flag: "--versions",
+        summary:
+          "列出 L0 基本信息的版本链（get_user_basic_info），用于找出覆盖了正确值的错误版本。",
+      },
+      USER_ID_OPTION,
+      JSON_OPTION,
+    ],
+    exit_codes: RUNTIME_EXITS,
+  },
+  {
+    name: "settings",
+    summary:
+      "查看或切换记忆模式（lite / pro / ultra）。lite 不调用模型，只在原始对话里检索；pro 由模型提取记忆（System 1）；ultra 在 pro 之上每晚自动进化（System 2，使用知识图谱）。任何模式都保留全部原始对话。切换不迁移数据：已处理的不会重做，排队中的和新的按新模式处理；需 --yes，不加时只显示计划。环境变量 ONE_MEMORY_GRAPH_STORE 会强制 ultra。settings features 查看或开关 SDK 的可选检索与提取功能（全局设置，默认全部保持 SDK 默认值）：graph_rerank 图谱重排（仅 ultra）、same_series_timeline 同系列时间线（需 graph_rerank）、episode_recall 片段召回（pro / ultra，需先开 session_episodes 写入片段）、session_episodes 片段写入、episode_summary 片段摘要（每个片段一次模型调用）、intention 意图提取（默认开）、few_shot 提取示例；set 需 --yes，不加时只显示计划；reset 恢复默认值。",
+    usage:
+      "memhub settings mode [lite|pro|ultra] [--yes] [--json]\n" +
+      "            memhub settings features [--json]\n" +
+      "            memhub settings features set <name>=<true|false> [...] [--yes] [--json]\n" +
+      "            memhub settings features reset [<name> ...] [--json]",
+    options: [
+      {
+        flag: "mode [lite|pro|ultra]",
+        summary: "不带模式只显示当前模式；带模式时先显示切换计划。",
+      },
+      {
+        flag: "features set <name>=<true|false>",
+        summary:
+          "开关一个或多个 SDK 功能；功能在当前模式下不生效时会注明原因（例如 graph_rerank 需要 ultra）。",
+      },
+      {
+        flag: "features reset [<name> ...]",
+        summary: "把指定功能（不写则全部）恢复为 SDK 默认值。",
+      },
+      CONFIRM_OPTION,
+      JSON_OPTION,
+    ],
+    exit_codes: RUNTIME_EXITS,
+  },
+  {
+    name: "evolve",
+    summary:
+      "立即为一个用户运行记忆进化（System 2，仅 ultra 模式）。在后台运行并返回运行 ID；--wait 等待结果。同一用户同时只运行一次；每晚的自动进化由 Runtime 按设置安排。",
+    usage:
+      "memhub evolve [--user-id <id>] [--agent-id <id>] [--max-batches <n>] [--wait] [--json]",
+    options: [
+      USER_ID_OPTION,
+      {
+        flag: "--agent-id <id>",
+        summary:
+          "进化哪个 agent 的记忆，默认 all-apps（MemHub 写入记忆用的 agent）。",
+      },
+      {
+        flag: "--max-batches <n>",
+        summary: "每轮最多处理的批次数（1–1000），默认由 SDK 决定。",
+      },
+      { flag: "--wait", summary: "等待这次进化结束再返回。" },
+      JSON_OPTION,
+    ],
     exit_codes: RUNTIME_EXITS,
   },
   {
@@ -475,7 +551,7 @@ export const COMMANDS: readonly CommandDescriptor[] = [
     summary:
       "从原始对话与已沉淀记忆统一检索上下文；本地只读，不调用模型、不触发补索引。文本输出标明每条是「原始对话摘录」还是「记忆」（JSON 里是 kind: raw|memory）。",
     usage:
-      "memhub recall <query> [--kind <raw|memory>] [--source <id>] [--limit <n>] [--character-budget <n>] [--timeout-ms <n>] [--json]\n            memhub recall --query-file <path> [...]\n            memhub recall --stdin [...]",
+      "memhub recall <query> [--kind <raw|memory>] [--source <id>] [--limit <n>] [--character-budget <n>] [--timeout-ms <n>] [--user-id <id>] [--json]\n            memhub recall --query-file <path> [...]\n            memhub recall --stdin [...]",
     options: [
       {
         flag: "--query-file <path>",
@@ -502,6 +578,7 @@ export const COMMANDS: readonly CommandDescriptor[] = [
         flag: "--timeout-ms <n>",
         summary: "Core 查询时限 50..30000 毫秒，默认 2000。",
       },
+      USER_ID_OPTION,
       JSON_OPTION,
     ],
     exit_codes: RUNTIME_EXITS,
@@ -650,7 +727,7 @@ export const COMMANDS: readonly CommandDescriptor[] = [
     summary:
       "提交一段 OpenAI 格式的对话（messages）到写入队列，异步加工成记忆；或按 request id 查询状态。request id 是幂等键：不给时自动生成并输出，重试请带上同一个。新增一条结论式记忆请用 memories add。",
     usage:
-      "memhub write submit --messages-file <path> --session <id> [--source <id>] [--request-id <id>] [--memory-at <RFC3339>] [--live] [--json]\n            memhub write status <request-id> [--json]",
+      "memhub write submit --messages-file <path> --session <id> [--source <id>] [--request-id <id>] [--memory-at <RFC3339>] [--live] [--user-id <id>] [--json]\n            memhub write status <request-id> [--user-id <id>] [--json]",
     options: [
       {
         flag: "--messages-file <path>",
@@ -666,6 +743,7 @@ export const COMMANDS: readonly CommandDescriptor[] = [
       },
       { flag: "--memory-at <RFC3339>", summary: "对话发生时间，必须带时区。" },
       { flag: "--live", summary: "优先处理。" },
+      USER_ID_OPTION,
       JSON_OPTION,
     ],
     exit_codes: RUNTIME_EXITS,

memhub/apps/cli/src/command-contract.test.ts

@@ -76,6 +76,8 @@ const COMMANDS = [
   "models",
   "memories",
   "profile",
+  "settings",
+  "evolve",
   "graph-width",
   "host",
   "recall",

memhub/apps/cli/src/context-search.ts

@@ -10,6 +10,8 @@ export interface ContextSearchInput {
   limit: number;
   character_budget: number;
   timeout_ms: number;
+  /** Whose memories to search. Omitted for the default user (`local`). */
+  user_id?: string;
 }
 
 /** The CLI's JSON document mirrors Core's SDK-native context schema v2. */

memhub/apps/cli/src/core-client.ts

@@ -12,11 +12,19 @@ const READ_ONLY_PATHS = [
   "/api/v1/live-sync",
   "/api/v1/memory-processing",
   "/api/v1/model-profiles",
+  "/api/v1/source-grants",
   "/api/v1/sources",
   "/api/v2/home-summary",
   "/api/v2/memory-library",
   "/api/v2/user-profile",
+  "/api/v2/user-profile/versions",
+  "/api/v1/settings/memory-mode",
+  "/api/v1/settings/memory-features",
+  "/api/v2/evolve",
 ] as const;
+/** One evolve run's receipt (`memhub evolve --wait`). */
+const EVOLVE_RUN =
+  /^\/api\/v2\/evolve\/[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}$/;
 const SDK_MEMORY_DETAIL = /^\/api\/v2\/memories\/[A-Za-z0-9][\w.-]{0,127}$/;
 /** Status of one `POST /api/v2/memories` addition, by its request id. */
 const MEMORY_ADDITION =
@@ -219,6 +227,7 @@ export function isReadOnlyPath(path: string): boolean {
   if (IMPORT_JOB.test(path)) return true;
   if (OPERATION_RECEIPT.test(path)) return true;
   if (MEMORY_ADDITION.test(path)) return true;
+  if (EVOLVE_RUN.test(path)) return true;
   if (!SDK_MEMORY_DETAIL.test(path)) return false;
   const segment = path.slice("/api/v2/memories/".length);
   return !(RESERVED_MEMORY_SEGMENTS as readonly string[]).includes(segment);
@@ -310,6 +319,8 @@ export async function submitMemoryWrite(
     messages: readonly OpenAIMessage[];
     memory_at?: string;
     priority?: "live" | "normal";
+    /** Whose memory this is. Omitted for the default user (`local`). */
+    user_id?: string;
   },
 ): Promise<MemoryWriteOutcome> {
   let response: { status: number; data: unknown };
@@ -327,6 +338,9 @@ export async function submitMemoryWrite(
         ...(request.priority === undefined
           ? {}
           : { priority: request.priority }),
+        ...(request.user_id === undefined || request.user_id === "local"
+          ? {}
+          : { user_id: request.user_id }),
       },
       {
         headers: { Authorization: `Bearer ${attachment.apiToken}` },
@@ -469,11 +483,16 @@ export async function deleteModelProfile(
 export async function readMemoryWrite(
   attachment: RuntimeAttachmentLike,
   requestId: string,
+  /** A receipt belongs to the user that submitted it; omit for `local`. */
+  userId?: string,
 ): Promise<MemoryWriteReceipt | null> {
   try {
     const response = await axios.get<MemoryWriteReceipt>(
       `${baseUrlOf(attachment)}/api/v2/memory-writes/${encodeURIComponent(requestId)}`,
       {
+        ...(userId === undefined || userId === "local"
+          ? {}
+          : { params: { user_id: userId } }),
         headers: { Authorization: `Bearer ${attachment.apiToken}` },
         timeout: 10_000,
         proxy: false,

memhub/apps/cli/src/core-operations.ts

@@ -6,9 +6,10 @@ const WRITE_PATHS = [
 ] as const;
 /** `POST /api/v2/memories` adds one memory (the app's 添加记忆). */
 const MEMORY_ADD_PATH = "/api/v2/memories";
-/** One SDK memory; delete names its namespace in the query string. */
+/** One SDK memory; delete names its namespace, and a non-default user, in
+ * the query string. */
 const SDK_MEMORY_PATH =
-  /^\/api\/v2\/memories\/[A-Za-z0-9][\w.-]{0,127}(?:\?kind=(?:chat|experience|fact)(?:&variant=(?:working|coding))?)?$/;
+  /^\/api\/v2\/memories\/[A-Za-z0-9][\w.-]{0,127}(?:\?(?:kind=(?:chat|experience|fact)(?:&variant=(?:working|coding))?(?:&user_id=[A-Za-z0-9._-]{1,64})?|user_id=[A-Za-z0-9._-]{1,64}))?$/;
 const MODEL_PROBE_PATH =
   /^\/api\/v1\/model-profiles\/[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}\/probe$/;
 const MODEL_PROFILE_PATH =
@@ -25,6 +26,11 @@ const SOURCE_IMPORT_JOB_PATH = "/api/v1/import-jobs";
 const IMPORT_JOB_ACTION_PATH =
   /^\/api\/v1\/import-jobs\/[A-Za-z0-9][\w.-]{0,127}\/(?:pause|resume|cancel)$/;
 const GRAPH_WIDTH_PATH = "/api/v1/graph-width";
+/** `memhub evolve` starts a run; `memhub settings mode <m> --yes` sets the mode. */
+const EVOLVE_PATH = "/api/v2/evolve";
+const MEMORY_MODE_PATH = "/api/v1/settings/memory-mode";
+/** `memhub settings features set|reset` sets the SDK features. */
+const MEMORY_FEATURES_PATH = "/api/v1/settings/memory-features";
 const ONBOARDING_COMPLETE_PATH = "/api/v1/onboarding/complete";
 const MAX_BYTES = 4 * 1024 * 1024;
 
@@ -96,6 +102,7 @@ export function createCoreOperations(attachment: {
         !SOURCE_AUTHORIZE_PATH.test(path) &&
         !SOURCE_REVOKE_PATH.test(path) &&
         path !== GRAPH_WIDTH_PATH &&
+        path !== EVOLVE_PATH &&
         path !== ONBOARDING_COMPLETE_PATH &&
         !MODEL_PROFILE_ACTIVATE.test(path)
       ) {
@@ -131,7 +138,11 @@ export function createCoreOperations(attachment: {
       body: Record<string, unknown>,
       signal?: AbortSignal,
     ) {
-      if (!SDK_MEMORY_PATH.test(path) || path.includes("?")) {
+      if (
+        path !== MEMORY_MODE_PATH &&
+        path !== MEMORY_FEATURES_PATH &&
+        (!SDK_MEMORY_PATH.test(path) || path.includes("?"))
+      ) {
         throw new Error("core_operation_path_not_allowed");
       }
       let response: { status: number; data: unknown };

memhub/apps/cli/src/headless-flows.ts

@@ -110,9 +110,18 @@ export async function connectSourcesWithImport(
     sourceIds: readonly string[];
     confirmOverrides: boolean;
     requestId?: string;
+    /** Whose grant this is. Omitted for the default user (`local`). */
+    userId?: string;
+    /** An absolute root other than the source's candidate (a Hermes
+     * profile). Core reads its identity itself, so the source's candidate
+     * root id and revision are not sent for it. */
+    root?: string;
   },
 ): Promise<FlowDocument> {
   const sourceIds = input.sourceIds;
+  // A second user's grant is not a first run, and a custom root has no
+  // candidate the app showed: both are a plain connect.
+  const profile = input.root !== undefined || input.userId !== undefined;
   if (
     sourceIds.length < 1 ||
     sourceIds.length > 32 ||
@@ -134,9 +143,14 @@ export async function connectSourcesWithImport(
     return invalid("source_unavailable");
   }
   const items = selected.map(([candidate]) => candidate);
-  const needsOverride = items
-    .filter((item) => item.requires_override_confirmation)
-    .map((item) => item.source_id);
+  // A custom root is the caller's own choice, not the environment override
+  // the confirmation exists for.
+  const needsOverride =
+    input.root !== undefined
+      ? []
+      : items
+          .filter((item) => item.requires_override_confirmation)
+          .map((item) => item.source_id);
   if (needsOverride.length > 0 && !input.confirmOverrides) {
     return {
       schema_version: 1,
@@ -154,18 +168,22 @@ export async function connectSourcesWithImport(
     request = buildSourceImportRequest({
       sourceIds: [...sourceIds],
       sourceRoots: Object.fromEntries(
-        items.map((item) => [item.source_id, item.candidate_root_id]),
+        items.flatMap((item) =>
+          input.root === undefined
+            ? [[item.source_id, item.candidate_root_id]]
+            : [],
+        ),
       ),
       sourceCandidateRevisions: Object.fromEntries(
         items.flatMap((item) =>
-          item.candidate_revision === undefined
+          item.candidate_revision === undefined || input.root !== undefined
             ? []
             : [[item.source_id, item.candidate_revision]],
         ),
       ),
       confirmEnvironmentOverrides: needsOverride.length > 0,
       fullRescan: false,
-      onboarding: true,
+      onboarding: !profile,
       enableSync: true,
       requestId,
     });
@@ -175,7 +193,11 @@ export async function connectSourcesWithImport(
     }
     throw error;
   }
-  const result = await operations.post<unknown>(request.path, request.body);
+  const result = await operations.post<unknown>(request.path, {
+    ...request.body,
+    ...(input.userId === undefined ? {} : { user_id: input.userId }),
+    ...(input.root === undefined ? {} : { root: input.root }),
+  });
   if (result.state === "rejected") {
     return rejected(requestId, result.code);
   }

memhub/apps/cli/src/log-report.ts

@@ -107,7 +107,11 @@ export async function readLogReport(
   query: LogQuery,
 ): Promise<{
   events: DiagnosticEventV1[];
+  /** Lines that are not JSON: the only ones `memhub logs` calls damaged. */
   invalid_lines: number;
+  /** Whole JSON lines the reader leaves out, such as an older build's
+   * records. Nothing is wrong with them, so the text output says nothing. */
+  unrecognized_lines: number;
   failure?: "log_directory_unreadable";
 }> {
   // `readDiagnosticEvents` treats an unreadable logs path as "no events". That
@@ -119,6 +123,7 @@ export async function readLogReport(
     return {
       events: [],
       invalid_lines: 0,
+      unrecognized_lines: 0,
       failure: "log_directory_unreadable",
     };
   }
@@ -129,6 +134,7 @@ export async function readLogReport(
     return {
       events: [],
       invalid_lines: 0,
+      unrecognized_lines: 0,
       failure: "log_directory_unreadable",
     };
   }

memhub/apps/cli/src/memory-features-cli.test.ts

@@ -0,0 +1,318 @@
+import { beforeEach, describe, expect, it, vi } from "vitest";
+
+import { run } from "./cli.js";
+import { parseFeaturesCommand } from "./memory-features-flow.js";
+import { composeStatusReport, renderStatusText } from "./status-report.js";
+
+const transport = vi.hoisted(() => ({
+  get: vi.fn(),
+  post: vi.fn(),
+  put: vi.fn(),
+  patch: vi.fn(),
+  delete: vi.fn(),
+}));
+vi.mock("axios", () => ({
+  default: {
+    create: vi.fn(() => transport),
+    get: vi.fn(),
+    post: vi.fn(),
+    isAxiosError: () => false,
+  },
+}));
+
+const runtime = vi.hoisted(() => ({
+  inspectRuntimeStatus: vi.fn(),
+  discoverRuntimeAttachment: vi.fn(),
+}));
+vi.mock("@one-memory/runtime-control", async (importOriginal) => ({
+  ...(await importOriginal<typeof import("@one-memory/runtime-control")>()),
+  ...runtime,
+}));
+
+const DEFAULTS: Record<string, boolean> = {
+  graph_rerank: false,
+  same_series_timeline: false,
+  episode_recall: false,
+  session_episodes: false,
+  episode_summary: false,
+  intention: true,
+  few_shot: false,
+};
+
+function featuresState(
+  configured: Record<string, boolean> = {},
+  mode = "pro",
+  extra: Record<string, unknown> = {},
+) {
+  const nonDefault: Record<string, boolean> = {};
+  const features = Object.entries(DEFAULTS).map(([name, fallback]) => {
+    const enabled = configured[name] ?? fallback;
+    if (enabled !== fallback) nonDefault[name] = enabled;
+    const needsUltra = name === "graph_rerank";
+    const reason = !enabled
+      ? "disabled"
+      : needsUltra && mode !== "ultra"
+        ? "requires_ultra"
+        : null;
+    return {
+      name,
+      enabled,
+      configured: name in configured ? configured[name] : null,
+      default: fallback,
+      active: reason === null,
+      reason,
+      side: "read",
+      requires_mode: needsUltra ? "ultra" : "pro",
+      requires_feature: null,
+      cost: "none",
+      sdk_field: `recall.${name}`,
+      env: `MEMORY_${name.toUpperCase()}`,
+      description: name,
+    };
+  });
+  return { mode, features, non_default: nonDefault, ...extra };
+}
+
+beforeEach(() => {
+  vi.clearAllMocks();
+  for (const method of Object.values(transport)) method.mockReset();
+  runtime.inspectRuntimeStatus.mockResolvedValue({
+    state: "running",
+    code: "runtime_verified",
+    instanceId: "instance-1",
+    host: "127.0.0.1",
+    port: 8765,
+    version: "1.2.3",
+  });
+  runtime.discoverRuntimeAttachment.mockResolvedValue({
+    instanceId: "instance-1",
+    apiToken: "synthetic-runtime-token",
+    host: "127.0.0.1",
+    port: 8765,
+    servicePid: 4242,
+    health: {
+      status: "ok",
+      service: "one-memory-core",
+      version: "1.2.3",
+      instance_id: "instance-1",
+    },
+  });
+  transport.get.mockImplementation(async (url: string) =>
+    url === "/api/v1/settings/memory-features"
+      ? { status: 200, data: featuresState() }
+      : { status: 404, data: {} },
+  );
+});
+
+describe("memhub settings features", () => {
+  it("lists every feature with its value and whether it acts", async () => {
+    const result = await capturedRun(["settings", "features", "--json"]);
+    expect(result.code).toBe(0);
+    const json = JSON.parse(result.stdout);
+    expect(json.state).toBe("ok");
+    expect(json.features.features).toHaveLength(7);
+    expect(transport.put).not.toHaveBeenCalled();
+    const text = (await capturedRun(["settings", "features"])).stdout;
+    expect(text).toContain("graph_rerank：关（默认）");
+    expect(text).toContain("intention：开（默认）");
+    expect(text).toContain("先开启 session_episodes");
+  });
+
+  it("prints the plan and changes nothing without --yes", async () => {
+    const result = await capturedRun([
+      "settings",
+      "features",
+      "set",
+      "graph_rerank=true",
+      "episode_recall=true",
+    ]);
+    expect(result.code).toBe(0);
+    expect(transport.put).not.toHaveBeenCalled();
+    expect(result.stdout).toContain(
+      "memhub settings features set graph_rerank=true episode_recall=true --yes",
+    );
+  });
+
+  it("sets features with --yes and reports the ones inactive in this mode", async () => {
+    transport.put.mockResolvedValue({
+      status: 200,
+      data: featuresState({ graph_rerank: true, episode_recall: true }, "pro", {
+        changed: ["graph_rerank", "episode_recall"],
+      }),
+    });
+    const result = await capturedRun([
+      "settings",
+      "features",
+      "set",
+      "graph_rerank=true",
+      "episode_recall=true",
+      "--yes",
+    ]);
+    expect(result.code).toBe(0);
+    expect(transport.put).toHaveBeenCalledWith(
+      "/api/v1/settings/memory-features",
+      { graph_rerank: true, episode_recall: true },
+      undefined,
+    );
+    expect(result.stdout).toContain("已更新：graph_rerank、episode_recall");
+    expect(result.stdout).toContain(
+      "graph_rerank：开（已设置，当前模式下不生效（需要 ultra））",
+    );
+  });
+
+  it("does nothing when the values are already set", async () => {
+    transport.get.mockResolvedValue({
+      status: 200,
+      data: featuresState({ few_shot: true }),
+    });
+    const result = await capturedRun([
+      "settings",
+      "features",
+      "set",
+      "few_shot=true",
+      "--yes",
+    ]);
+    expect(result.code).toBe(0);
+    expect(result.stdout).toContain("没有需要改动的设置");
+    expect(transport.put).not.toHaveBeenCalled();
+  });
+
+  it("resets the named features, or every set one, to the default", async () => {
+    transport.get.mockResolvedValue({
+      status: 200,
+      data: featuresState({ few_shot: true, intention: false }),
+    });
+    transport.put.mockResolvedValue({
+      status: 200,
+      data: featuresState({}, "pro", { changed: ["intention", "few_shot"] }),
+    });
+    expect(
+      (await capturedRun(["settings", "features", "reset", "few_shot"])).code,
+    ).toBe(0);
+    expect(transport.put).toHaveBeenLastCalledWith(
+      "/api/v1/settings/memory-features",
+      { few_shot: null },
+      undefined,
+    );
+    const all = await capturedRun(["settings", "features", "reset", "--json"]);
+    expect(JSON.parse(all.stdout).state).toBe("committed");
+    expect(transport.put).toHaveBeenLastCalledWith(
+      "/api/v1/settings/memory-features",
+      { intention: null, few_shot: null },
+      undefined,
+    );
+  });
+
+  it("refuses unknown names and malformed values before the Runtime", async () => {
+    for (const bad of [
+      ["settings", "features", "set"],
+      ["settings", "features", "set", "local_rerank=true"],
+      ["settings", "features", "set", "graph_rerank=yes"],
+      ["settings", "features", "set", "graph_rerank"],
+      ["settings", "features", "set", "few_shot=true", "few_shot=false"],
+      ["settings", "features", "reset", "nope"],
+      ["settings", "features", "toggle"],
+    ]) {
+      const result = await capturedRun(bad);
+      expect(result.code, bad.join(" ")).toBe(2);
+    }
+    expect(transport.put).not.toHaveBeenCalled();
+  });
+
+  it("surfaces a 422 from the Runtime as rejected", async () => {
+    transport.put.mockResolvedValue({
+      status: 422,
+      data: { detail: "invalid_request" },
+    });
+    const result = await capturedRun([
+      "settings",
+      "features",
+      "set",
+      "few_shot=true",
+      "--yes",
+      "--json",
+    ]);
+    expect(result.code).toBe(1);
+    expect(JSON.parse(result.stdout).state).toBe("rejected");
+  });
+
+  it("parses the forms the help shows", () => {
+    expect(parseFeaturesCommand([])).toEqual({ kind: "show" });
+    expect(parseFeaturesCommand(["set", "intention=false", "--yes"])).toEqual({
+      kind: "set",
+      values: { intention: false },
+      confirmed: true,
+    });
+    expect(parseFeaturesCommand(["reset"])).toEqual({
+      kind: "reset",
+      names: [],
+    });
+  });
+});
+
+describe("status pipeline features", () => {
+  it("shows the features that differ from the defaults", async () => {
+    const processing = {
+      pending: 0,
+      fresh_pending: 0,
+      backlog_pending: 0,
+      outside_window_pending: 0,
+      running: 0,
+      waiting_for_model: 0,
+      completed: 3,
+      failed: 0,
+      blocked_stages: [],
+      pipeline: {
+        mode: "pro",
+        configured_mode: "pro",
+        forced_by_env: false,
+        last_evolve_at: null,
+        graph_build: { state: "idle", users: {} },
+        evolve: { mode: "pro", schedule: null, running: [], users: {} },
+        features: { episode_recall: true, intention: false },
+      },
+    };
+    const report = await composeStatusReport({
+      dataDir: "/tmp/isolated",
+      runtime: { state: "running", code: "ok", host: "127.0.0.1", port: 1 },
+      background: null,
+      reader: (async (path: string) =>
+        path === "/api/v1/memory-processing"
+          ? { state: "ok", data: processing }
+          : { state: "unavailable", code: "x" }) as never,
+    });
+    expect(report.pipeline?.features).toEqual({
+      episode_recall: true,
+      intention: false,
+    });
+    const text = renderStatusText(report, "MemHub");
+    expect(text).toContain("SDK 功能：episode_recall=开、intention=关");
+  });
+});
+
+async function capturedRun(arguments_: string[]) {
+  let stdout = "";
+  let stderr = "";
+  const stdoutSpy = vi
+    .spyOn(process.stdout, "write")
+    .mockImplementation((chunk) => {
+      stdout += String(chunk);
+      return true;
+    });
+  const stderrSpy = vi
+    .spyOn(process.stderr, "write")
+    .mockImplementation((chunk) => {
+      stderr += String(chunk);
+      return true;
+    });
+  try {
+    return {
+      code: await run(arguments_, { ONE_MEMORY_DATA_DIR: "/synthetic/data" }),
+      stdout,
+      stderr,
+    };
+  } finally {
+    stdoutSpy.mockRestore();
+    stderrSpy.mockRestore();
+  }
+}

memhub/apps/cli/src/memory-features-flow.ts

@@ -0,0 +1,273 @@
+/**
+ * `memhub settings features [set <name>=<true|false> ... | reset [<name> ...]]`.
+ *
+ * The SDK's optional retrieval and extraction features (graph rerank, episode
+ * recall, the episode write side, intention, few-shot). One global setting in
+ * the Runtime; an unset feature keeps the SDK default, so an install nobody
+ * touched behaves as before. `set` prints its plan first and needs `--yes`,
+ * like `settings mode`, because some features spend model calls; `reset`
+ * only returns features to their defaults and applies at once.
+ */
+
+import type { CoreReader } from "./core-client.js";
+import type { CorePutOperations } from "./core-operations.js";
+import { PRODUCT } from "./version.js";
+
+export const MEMORY_FEATURES = [
+  "graph_rerank",
+  "same_series_timeline",
+  "episode_recall",
+  "session_episodes",
+  "episode_summary",
+  "intention",
+  "few_shot",
+] as const;
+export type MemoryFeature = (typeof MEMORY_FEATURES)[number];
+
+const FEATURE_LABEL: Record<MemoryFeature, string> = {
+  graph_rerank: "图谱重排（沿知识图谱的边传播支持度；仅 ultra）",
+  same_series_timeline:
+    "同系列时间线（同一系列的记忆按时间放在一起；需图谱重排）",
+  episode_recall:
+    "片段召回（补入同一片段的记忆并多通道打分；pro / ultra，需片段写入）",
+  session_episodes: "片段写入（同一会话提取的记忆归入同一片段）",
+  episode_summary: "片段摘要（片段结束时总结一次；每个片段一次模型调用）",
+  intention: "意图提取（随事实一起提取用户的计划与意图）",
+  few_shot: "提取示例（在提取提示里附上示例；不加调用，输入更长）",
+};
+
+const REASON_TEXT: Record<string, string> = {
+  disabled: "未开启",
+  requires_pro: "lite 模式下不生效（需要 pro 或 ultra）",
+  requires_ultra: "当前模式下不生效（需要 ultra）",
+  requires_graph_rerank: "需要先开启并生效 graph_rerank",
+};
+
+export type FeaturesCommand =
+  | { kind: "show" }
+  | {
+      kind: "set";
+      values: Partial<Record<MemoryFeature, boolean>>;
+      confirmed: boolean;
+    }
+  | { kind: "reset"; names: MemoryFeature[] };
+
+function isFeature(value: string): value is MemoryFeature {
+  return (MEMORY_FEATURES as readonly string[]).includes(value);
+}
+
+/** Arguments after `settings features`; null = usage. */
+export function parseFeaturesCommand(
+  values: readonly string[],
+): FeaturesCommand | null {
+  const rest = values.filter((value) => value !== "--json");
+  if (rest.length === 0) return { kind: "show" };
+  const [verb, ...tail] = rest;
+  const yes = tail.filter((value) => value === "--yes").length;
+  if (yes > 1) return null;
+  const positional = tail.filter((value) => value !== "--yes");
+  if (verb === "set") {
+    if (positional.length === 0) return null;
+    const parsed: Partial<Record<MemoryFeature, boolean>> = {};
+    for (const item of positional) {
+      const match = /^([a-z_]+)=(true|false)$/.exec(item);
+      if (!match || !isFeature(match[1]) || match[1] in parsed) return null;
+      parsed[match[1]] = match[2] === "true";
+    }
+    return { kind: "set", values: parsed, confirmed: yes === 1 };
+  }
+  if (verb === "reset") {
+    if (new Set(positional).size !== positional.length) return null;
+    if (!positional.every(isFeature)) return null;
+    return { kind: "reset", names: positional as MemoryFeature[] };
+  }
+  return null;
+}
+
+export interface FeatureState {
+  name: MemoryFeature;
+  enabled: boolean;
+  configured: boolean | null;
+  default: boolean;
+  active: boolean;
+  reason: string | null;
+  cost: string;
+  requires_mode: string;
+  description: string;
+}
+
+export interface FeaturesState {
+  mode: string;
+  features: FeatureState[];
+  non_default: Partial<Record<MemoryFeature, boolean>>;
+  changed?: MemoryFeature[];
+}
+
+function featuresState(value: unknown): FeaturesState | null {
+  if (!value || typeof value !== "object") return null;
+  const item = value as Record<string, unknown>;
+  if (
+    typeof item.mode !== "string" ||
+    !Array.isArray(item.features) ||
+    !item.non_default ||
+    typeof item.non_default !== "object"
+  ) {
+    return null;
+  }
+  for (const feature of item.features) {
+    if (
+      !feature ||
+      typeof feature !== "object" ||
+      typeof (feature as FeatureState).name !== "string" ||
+      typeof (feature as FeatureState).enabled !== "boolean" ||
+      typeof (feature as FeatureState).active !== "boolean"
+    ) {
+      return null;
+    }
+  }
+  return item as unknown as FeaturesState;
+}
+
+export type FeaturesDocument = {
+  schema_version: 1;
+  product: string;
+  command: "settings features";
+} & (
+  | { state: "ok"; features: FeaturesState }
+  | {
+      state: "preview";
+      features: FeaturesState;
+      target: Partial<Record<MemoryFeature, boolean>>;
+      applied: false;
+      next_step: string;
+    }
+  | { state: "unchanged"; features: FeaturesState }
+  | { state: "committed"; features: FeaturesState; changed: MemoryFeature[] }
+  | { state: "rejected" | "unknown" | "unavailable"; code: string }
+);
+
+const BASE = {
+  schema_version: 1 as const,
+  product: PRODUCT,
+  command: "settings features" as const,
+};
+
+const PATH = "/api/v1/settings/memory-features";
+
+export async function runFeaturesCommand(
+  command: FeaturesCommand,
+  reader: CoreReader,
+  operations: CorePutOperations,
+): Promise<FeaturesDocument> {
+  const read = await reader<unknown>(PATH);
+  if (read.state !== "ok") {
+    return { ...BASE, state: "unavailable", code: read.code };
+  }
+  const current = featuresState(read.data);
+  if (current === null) {
+    return { ...BASE, state: "unavailable", code: "response_invalid" };
+  }
+  if (command.kind === "show") {
+    return { ...BASE, state: "ok", features: current };
+  }
+  let body: Record<string, unknown>;
+  if (command.kind === "set") {
+    const configured = new Map(
+      current.features.map((feature) => [feature.name, feature.configured]),
+    );
+    const differs = Object.entries(command.values).some(
+      ([name, value]) => configured.get(name as MemoryFeature) !== value,
+    );
+    if (!differs) return { ...BASE, state: "unchanged", features: current };
+    if (!command.confirmed) {
+      const pairs = Object.entries(command.values)
+        .map(([name, value]) => `${name}=${value}`)
+        .join(" ");
+      return {
+        ...BASE,
+        state: "preview",
+        features: current,
+        target: command.values,
+        applied: false,
+        next_step: `${PRODUCT} settings features set ${pairs} --yes`,
+      };
+    }
+    body = { ...command.values };
+  } else {
+    const names =
+      command.names.length > 0
+        ? command.names
+        : current.features
+            .filter((feature) => feature.configured !== null)
+            .map((feature) => feature.name);
+    if (
+      names.every(
+        (name) =>
+          current.features.find((feature) => feature.name === name)
+            ?.configured === null,
+      )
+    ) {
+      return { ...BASE, state: "unchanged", features: current };
+    }
+    body = Object.fromEntries(names.map((name) => [name, null]));
+  }
+  const result = await operations.put<unknown>(PATH, body);
+  if (result.state !== "committed") {
+    return { ...BASE, state: result.state, code: result.code };
+  }
+  const changed = featuresState(result.data);
+  if (changed === null) {
+    return { ...BASE, state: "unknown", code: "response_invalid" };
+  }
+  return {
+    ...BASE,
+    state: "committed",
+    features: changed,
+    changed: changed.changed ?? [],
+  };
+}
+
+export function featuresExitCode(document: FeaturesDocument): number {
+  return ["ok", "preview", "unchanged", "committed"].includes(document.state)
+    ? 0
+    : 1;
+}
+
+function featureLine(feature: FeatureState): string {
+  const label = FEATURE_LABEL[feature.name] ?? feature.name;
+  const value = feature.enabled ? "开" : "关";
+  const origin = feature.configured === null ? "默认" : "已设置";
+  const activity =
+    feature.enabled && !feature.active && feature.reason
+      ? `，${REASON_TEXT[feature.reason] ?? feature.reason}`
+      : "";
+  return `${feature.name}：${value}（${origin}${activity}）— ${label}`;
+}
+
+export function renderFeaturesText(document: FeaturesDocument): string {
+  if ("code" in document) {
+    return `SDK 功能未能${document.state === "unavailable" ? "读取" : "修改"}：${document.code}\n`;
+  }
+  const state = document.features;
+  const lines = [`当前模式：${state.mode}`];
+  for (const feature of state.features) lines.push(featureLine(feature));
+  lines.push(
+    "片段召回（episode_recall）只对带片段的记忆有用：先开启 session_episodes，之后提取的记忆才会带片段。",
+  );
+  if (document.state === "unchanged") lines.push("没有需要改动的设置。");
+  if (document.state === "preview") {
+    lines.push("将设置：");
+    for (const [name, value] of Object.entries(document.target)) {
+      lines.push(`- ${name} = ${value ? "开" : "关"}`);
+    }
+    lines.push(`未执行任何改动。确认后运行：${document.next_step}`);
+  }
+  if (document.state === "committed") {
+    lines.push(
+      document.changed.length > 0
+        ? `已更新：${document.changed.join("、")}。正在进行的写入按原设置完成，之后的写入和检索使用新设置。`
+        : "已保存，生效的值没有变化。",
+    );
+  }
+  return `${lines.join("\n")}\n`;
+}

memhub/apps/cli/src/memory-mode-cli.test.ts

@@ -0,0 +1,309 @@
+import { beforeEach, describe, expect, it, vi } from "vitest";
+
+import { run } from "./cli.js";
+import { modeEffects, parseSettingsCommand } from "./memory-mode-flow.js";
+import { composeStatusReport, renderStatusText } from "./status-report.js";
+
+const transport = vi.hoisted(() => ({
+  get: vi.fn(),
+  post: vi.fn(),
+  put: vi.fn(),
+  patch: vi.fn(),
+  delete: vi.fn(),
+}));
+vi.mock("axios", () => ({
+  default: {
+    create: vi.fn(() => transport),
+    get: vi.fn(),
+    post: vi.fn(),
+    isAxiosError: () => false,
+  },
+}));
+
+const runtime = vi.hoisted(() => ({
+  inspectRuntimeStatus: vi.fn(),
+  discoverRuntimeAttachment: vi.fn(),
+}));
+vi.mock("@one-memory/runtime-control", async (importOriginal) => ({
+  ...(await importOriginal<typeof import("@one-memory/runtime-control")>()),
+  ...runtime,
+}));
+
+const RUN_ID = "11111111-2222-4333-8444-555555555555";
+
+function modeState(mode: string, configured = mode) {
+  return {
+    mode,
+    configured_mode: configured,
+    forced_by_env: false,
+    modes: ["lite", "pro", "ultra"],
+    graph_build: { state: "idle", users: {} },
+  };
+}
+
+function receipt(state: string, extra: Record<string, unknown> = {}) {
+  return {
+    run_id: RUN_ID,
+    user_id: "wife",
+    agent_id: "all-apps",
+    state,
+    trigger: "manual",
+    started_at: "2026-11-01T03:00:00Z",
+    rounds: state === "running" ? 0 : 2,
+    tasks_processed: state === "running" ? 0 : 5,
+    batch_failures: 0,
+    capped: false,
+    elapsed_ms: 10,
+    results: { schemas: ["never printed"] },
+    ...extra,
+  };
+}
+
+beforeEach(() => {
+  vi.clearAllMocks();
+  for (const method of Object.values(transport)) method.mockReset();
+  runtime.inspectRuntimeStatus.mockResolvedValue({
+    state: "running",
+    code: "runtime_verified",
+    instanceId: "instance-1",
+    host: "127.0.0.1",
+    port: 8765,
+    version: "1.2.3",
+  });
+  runtime.discoverRuntimeAttachment.mockResolvedValue({
+    instanceId: "instance-1",
+    apiToken: "synthetic-runtime-token",
+    host: "127.0.0.1",
+    port: 8765,
+    servicePid: 4242,
+    health: {
+      status: "ok",
+      service: "one-memory-core",
+      version: "1.2.3",
+      instance_id: "instance-1",
+    },
+  });
+  transport.get.mockImplementation(async (url: string) =>
+    url === "/api/v1/settings/memory-mode"
+      ? { status: 200, data: modeState("pro") }
+      : { status: 404, data: {} },
+  );
+});
+
+describe("memhub settings mode", () => {
+  it("shows the current mode without changing it", async () => {
+    const result = await capturedRun(["settings", "mode", "--json"]);
+    expect(result.code).toBe(0);
+    expect(JSON.parse(result.stdout).mode.mode).toBe("pro");
+    expect(transport.put).not.toHaveBeenCalled();
+  });
+
+  it("prints the plan and changes nothing without --yes", async () => {
+    const result = await capturedRun(["settings", "mode", "lite"]);
+    expect(result.code).toBe(0);
+    expect(transport.put).not.toHaveBeenCalled();
+    expect(result.stdout).toContain("更高模式生成的记忆仍保存着");
+    expect(result.stdout).toContain("memhub settings mode lite --yes");
+    const json = JSON.parse(
+      (await capturedRun(["settings", "mode", "ultra", "--json"])).stdout,
+    );
+    expect(json.state).toBe("preview");
+    expect(json.applied).toBe(false);
+    expect(json.effects).toEqual(["existing_memories_added_to_graph"]);
+  });
+
+  it("changes the mode with --yes and reports what changed", async () => {
+    transport.put.mockResolvedValue({
+      status: 200,
+      data: {
+        ...modeState("lite"),
+        previous_mode: "pro",
+        effects: ["higher_mode_memories_hidden"],
+      },
+    });
+    const result = await capturedRun(["settings", "mode", "lite", "--yes"]);
+    expect(result.code).toBe(0);
+    expect(transport.put).toHaveBeenCalledWith(
+      "/api/v1/settings/memory-mode",
+      { mode: "lite" },
+      undefined,
+    );
+    expect(result.stdout).toContain("已切换");
+    expect(result.stdout).toContain("原始对话：全部保留");
+  });
+
+  it("does nothing for the mode already set, and refuses bad input", async () => {
+    const same = await capturedRun(["settings", "mode", "pro", "--yes"]);
+    expect(same.code).toBe(0);
+    expect(same.stdout).toContain("已经是这个模式");
+    expect(transport.put).not.toHaveBeenCalled();
+    for (const bad of [
+      ["settings"],
+      ["settings", "mode", "max"],
+      ["settings", "mode", "--yes"],
+      ["settings", "mode", "lite", "pro"],
+      ["settings", "language", "en"],
+    ]) {
+      expect((await capturedRun(bad)).code, bad.join(" ")).toBe(2);
+    }
+  });
+
+  it("plans the same effects Core reports", () => {
+    expect(modeEffects("lite", "pro")).toEqual([
+      "earlier_raw_not_extracted",
+      "lite_search_stops",
+    ]);
+    expect(modeEffects("ultra", "lite")).toEqual([
+      "higher_mode_memories_hidden",
+      "evolve_stopped",
+    ]);
+    expect(modeEffects("pro", "pro")).toEqual([]);
+    expect(parseSettingsCommand(["mode", "ultra", "--yes"])).toEqual({
+      kind: "set",
+      mode: "ultra",
+      confirmed: true,
+    });
+  });
+});
+
+describe("memhub evolve", () => {
+  it("starts a run for the named user and returns at once", async () => {
+    transport.post.mockResolvedValue({ status: 202, data: receipt("running") });
+    const result = await capturedRun([
+      "evolve",
+      "--user-id",
+      "wife",
+      "--agent-id",
+      "all-apps",
+      "--max-batches",
+      "3",
+      "--json",
+    ]);
+    expect(result.code).toBe(0);
+    expect(transport.post).toHaveBeenCalledWith(
+      "/api/v2/evolve",
+      { user_id: "wife", agent_id: "all-apps", max_batches: 3 },
+      undefined,
+    );
+    const document = JSON.parse(result.stdout);
+    expect(document.state).toBe("running");
+    expect(document.run.results).toBeUndefined();
+  });
+
+  it("follows the run with --wait until it ends", async () => {
+    transport.post.mockResolvedValue({ status: 202, data: receipt("running") });
+    const reads = [receipt("running"), receipt("succeeded")];
+    transport.get.mockImplementation(async (url: string) =>
+      url === `/api/v2/evolve/${RUN_ID}`
+        ? { status: 200, data: reads.shift() }
+        : { status: 404, data: {} },
+    );
+    const result = await capturedRun(["evolve", "--wait"], {
+      sleep: async () => undefined,
+    });
+    expect(result.code).toBe(0);
+    expect(transport.post).toHaveBeenCalledWith(
+      "/api/v2/evolve",
+      {},
+      undefined,
+    );
+    expect(result.stdout).toContain("进化完成");
+    expect(result.stdout).toContain("2 轮");
+  });
+
+  it("names the mode when Core refuses outside ultra", async () => {
+    transport.post.mockResolvedValue({
+      status: 409,
+      data: { detail: "memory_mode_not_ultra" },
+    });
+    const result = await capturedRun(["evolve"]);
+    expect(result.code).toBe(1);
+    expect(result.stdout).toContain("memhub settings mode ultra --yes");
+    expect((await capturedRun(["evolve", "--max-batches", "0"])).code).toBe(2);
+    expect((await capturedRun(["evolve", "--now"])).code).toBe(2);
+  });
+});
+
+describe("status pipeline", () => {
+  it("reports pipeline.mode and the last evolve from memory-processing", async () => {
+    const processing = {
+      pending: 0,
+      fresh_pending: 0,
+      backlog_pending: 0,
+      outside_window_pending: 0,
+      running: 0,
+      waiting_for_model: 0,
+      completed: 3,
+      failed: 0,
+      blocked_stages: [],
+      pipeline: {
+        mode: "ultra",
+        configured_mode: "ultra",
+        forced_by_env: false,
+        last_evolve_at: "2026-11-01T03:00:05+08:00",
+        graph_build: { state: "succeeded", users: {} },
+        evolve: {
+          mode: "ultra",
+          schedule: { enabled: true, time: "03:00" },
+          running: [],
+          users: {
+            local: {
+              last_run_at: "2026-11-01T03:00:05+08:00",
+              last_result: { state: "succeeded", tasks_processed: 4 },
+              last_error: null,
+            },
+          },
+        },
+      },
+    };
+    const report = await composeStatusReport({
+      dataDir: "/tmp/isolated",
+      runtime: { state: "running", code: "ok", host: "127.0.0.1", port: 1 },
+      background: null,
+      reader: (async (path: string) =>
+        path === "/api/v1/memory-processing"
+          ? { state: "ok", data: processing }
+          : { state: "unavailable", code: "x" }) as never,
+    });
+    expect(report.pipeline?.mode).toBe("ultra");
+    expect(report.pipeline?.last_evolve_at).toBe("2026-11-01T03:00:05+08:00");
+    expect(report.pipeline?.evolve_users.local?.state).toBe("succeeded");
+    const text = renderStatusText(report, "MemHub");
+    expect(text).toContain("记忆模式：ultra");
+    expect(text).toContain("每日 03:00 自动进化");
+  });
+});
+
+async function capturedRun(
+  arguments_: string[],
+  controls: Parameters<typeof run>[2] = {},
+) {
+  let stdout = "";
+  let stderr = "";
+  const stdoutSpy = vi
+    .spyOn(process.stdout, "write")
+    .mockImplementation((chunk) => {
+      stdout += String(chunk);
+      return true;
+    });
+  const stderrSpy = vi
+    .spyOn(process.stderr, "write")
+    .mockImplementation((chunk) => {
+      stderr += String(chunk);
+      return true;
+    });
+  try {
+    return {
+      code: await run(
+        arguments_,
+        { ONE_MEMORY_DATA_DIR: "/synthetic/data" },
+        controls,
+      ),
+      stdout,
+      stderr,
+    };
+  } finally {
+    stdoutSpy.mockRestore();
+    stderrSpy.mockRestore();
+  }
+}

memhub/apps/cli/src/memory-mode-flow.ts

@@ -0,0 +1,372 @@
+/**
+ * `memhub settings mode` and `memhub evolve`.
+ *
+ * The mode (lite / pro / ultra) decides what the SDK does with every
+ * conversation, and switching it migrates nothing: what one mode made is not
+ * shown by a lower one, and raw kept by lite is not extracted until it is
+ * imported again. So a change prints its plan first and needs `--yes`, like
+ * every other command whose effect a person cannot simply take back.
+ *
+ * `memhub evolve` starts System 2 (ultra only) and, with `--wait`, follows
+ * the run until it ends.
+ */
+
+import type { CoreReader } from "./core-client.js";
+import type { CoreOperations, CorePutOperations } from "./core-operations.js";
+import { PRODUCT } from "./version.js";
+
+export const MEMORY_MODES = ["lite", "pro", "ultra"] as const;
+export type MemoryMode = (typeof MEMORY_MODES)[number];
+
+export type ModeEffect =
+  | "earlier_raw_not_extracted"
+  | "higher_mode_memories_hidden"
+  | "existing_memories_added_to_graph"
+  | "evolve_stopped"
+  | "lite_search_stops"
+  | "graph_backfill_already_running";
+
+const MODE_SUMMARY: Record<MemoryMode, string> = {
+  lite: "lite：不调用模型，保留全部原始对话，只在原始对话里检索",
+  pro: "pro：模型从对话中提取记忆（System 1）",
+  ultra: "ultra：pro + 每晚自动进化（System 2，使用知识图谱）",
+};
+
+const EFFECT_TEXT: Record<ModeEffect, string> = {
+  earlier_raw_not_extracted:
+    "lite 下已处理完的对话不会再被提取；只有尚未处理的和之后的新对话会由模型整理成记忆，检索立即改为记忆检索",
+  higher_mode_memories_hidden:
+    "更高模式生成的记忆仍保存着，但在这个模式下不会显示",
+  existing_memories_added_to_graph:
+    "已有记忆会在后台加入知识图谱（每个用户一次）",
+  evolve_stopped: "每晚的自动进化（System 2）停止",
+  lite_search_stops: "lite 只在原始对话里的检索立即停止，改为检索记忆",
+  graph_backfill_already_running: "知识图谱的补建已在进行中",
+};
+
+/** The same rule Core applies (`memory_mode_effects`), for the plan. */
+export function modeEffects(
+  previous: MemoryMode,
+  next: MemoryMode,
+): ModeEffect[] {
+  if (previous === next) return [];
+  const rank: Record<MemoryMode, number> = { lite: 0, pro: 1, ultra: 2 };
+  const effects: ModeEffect[] = [];
+  if (previous === "lite") {
+    effects.push("earlier_raw_not_extracted", "lite_search_stops");
+  }
+  if (rank[next] < rank[previous]) effects.push("higher_mode_memories_hidden");
+  if (previous === "ultra") effects.push("evolve_stopped");
+  if (next === "ultra") effects.push("existing_memories_added_to_graph");
+  return effects;
+}
+
+export type ModeCommand =
+  | { kind: "show" }
+  | { kind: "set"; mode: MemoryMode; confirmed: boolean };
+
+/** `settings mode [lite|pro|ultra] [--yes]` after `settings`; null = usage. */
+export function parseSettingsCommand(
+  values: readonly string[],
+): ModeCommand | null {
+  const rest = values.filter((value) => value !== "--json");
+  if (rest[0] !== "mode") return null;
+  const tail = rest.slice(1);
+  const confirmed = tail.includes("--yes");
+  const positional = tail.filter((value) => value !== "--yes");
+  if (tail.filter((value) => value === "--yes").length > 1) return null;
+  if (positional.length === 0) return confirmed ? null : { kind: "show" };
+  if (positional.length !== 1) return null;
+  const mode = positional[0] as MemoryMode;
+  if (!(MEMORY_MODES as readonly string[]).includes(mode)) return null;
+  return { kind: "set", mode, confirmed };
+}
+
+export interface ModeState {
+  mode: MemoryMode;
+  configured_mode: MemoryMode;
+  forced_by_env: boolean;
+  graph_build?: { state?: string } | null;
+  previous_mode?: MemoryMode;
+  effects?: ModeEffect[];
+}
+
+function modeState(value: unknown): ModeState | null {
+  if (!value || typeof value !== "object") return null;
+  const item = value as Record<string, unknown>;
+  if (
+    !(MEMORY_MODES as readonly unknown[]).includes(item.mode) ||
+    !(MEMORY_MODES as readonly unknown[]).includes(item.configured_mode) ||
+    typeof item.forced_by_env !== "boolean"
+  ) {
+    return null;
+  }
+  return item as unknown as ModeState;
+}
+
+export type ModeDocument = {
+  schema_version: 1;
+  product: string;
+  command: "settings mode";
+} & (
+  | { state: "ok"; mode: ModeState }
+  | {
+      state: "preview";
+      mode: ModeState;
+      target: MemoryMode;
+      effects: ModeEffect[];
+      applied: false;
+      next_step: string;
+    }
+  | { state: "unchanged"; mode: ModeState }
+  | { state: "committed"; mode: ModeState; effects: ModeEffect[] }
+  | { state: "rejected" | "unknown" | "unavailable"; code: string }
+);
+
+const BASE = {
+  schema_version: 1 as const,
+  product: PRODUCT,
+  command: "settings mode" as const,
+};
+
+export async function runModeCommand(
+  command: ModeCommand,
+  reader: CoreReader,
+  operations: CorePutOperations,
+): Promise<ModeDocument> {
+  const read = await reader<unknown>("/api/v1/settings/memory-mode");
+  if (read.state !== "ok") {
+    return { ...BASE, state: "unavailable", code: read.code };
+  }
+  const current = modeState(read.data);
+  if (current === null) {
+    return { ...BASE, state: "unavailable", code: "response_invalid" };
+  }
+  if (command.kind === "show") return { ...BASE, state: "ok", mode: current };
+  if (current.configured_mode === command.mode) {
+    return { ...BASE, state: "unchanged", mode: current };
+  }
+  if (!command.confirmed) {
+    return {
+      ...BASE,
+      state: "preview",
+      mode: current,
+      target: command.mode,
+      effects: modeEffects(current.mode, command.mode),
+      applied: false,
+      next_step: `${PRODUCT} settings mode ${command.mode} --yes`,
+    };
+  }
+  const result = await operations.put<unknown>("/api/v1/settings/memory-mode", {
+    mode: command.mode,
+  });
+  if (result.state !== "committed") {
+    return { ...BASE, state: result.state, code: result.code };
+  }
+  const changed = modeState(result.data);
+  if (changed === null) {
+    return { ...BASE, state: "unknown", code: "response_invalid" };
+  }
+  return {
+    ...BASE,
+    state: "committed",
+    mode: changed,
+    effects: changed.effects ?? [],
+  };
+}
+
+export function modeExitCode(document: ModeDocument): number {
+  return ["ok", "preview", "unchanged", "committed"].includes(document.state)
+    ? 0
+    : 1;
+}
+
+export function renderModeText(document: ModeDocument): string {
+  if ("code" in document) {
+    return `记忆模式未能${document.state === "unavailable" ? "读取" : "修改"}：${document.code}\n`;
+  }
+  const mode = document.mode;
+  const lines = [`当前模式：${MODE_SUMMARY[mode.mode]}`];
+  if (mode.forced_by_env) {
+    lines.push(
+      `设置里是 ${mode.configured_mode}；环境变量 ONE_MEMORY_GRAPH_STORE 强制使用 ultra。`,
+    );
+  }
+  lines.push("原始对话：全部保留（任何模式都不删除）");
+  if (document.state === "unchanged") {
+    lines.push("已经是这个模式，未做改动。");
+  }
+  if (document.state === "preview") {
+    lines.push(`将切换到：${MODE_SUMMARY[document.target]}`);
+    for (const effect of document.effects)
+      lines.push(`- ${EFFECT_TEXT[effect]}`);
+    lines.push(`未执行任何改动。确认后运行：${document.next_step}`);
+  }
+  if (document.state === "committed") {
+    lines.push("已切换。");
+    for (const effect of document.effects)
+      lines.push(`- ${EFFECT_TEXT[effect]}`);
+  }
+  return `${lines.join("\n")}\n`;
+}
+
+export type EvolveCommand = {
+  agentId?: string;
+  maxBatches?: number;
+  wait: boolean;
+};
+
+/** `evolve [--agent-id <id>] [--max-batches <n>] [--wait]`; null = usage. */
+export function parseEvolveCommand(
+  values: readonly string[],
+): EvolveCommand | null {
+  const parsed: EvolveCommand = { wait: false };
+  const rest = values.filter((value) => value !== "--json");
+  for (let index = 0; index < rest.length; index += 1) {
+    const flag = rest[index];
+    if (flag === "--wait" && !parsed.wait) {
+      parsed.wait = true;
+    } else if (flag === "--agent-id" && parsed.agentId === undefined) {
+      const value = rest[index + 1];
+      if (!value || !/^[A-Za-z0-9._:-]{1,128}$/.test(value)) return null;
+      parsed.agentId = value;
+      index += 1;
+    } else if (flag === "--max-batches" && parsed.maxBatches === undefined) {
+      const value = Number(rest[index + 1]);
+      if (!Number.isInteger(value) || value < 1 || value > 1000) return null;
+      parsed.maxBatches = value;
+      index += 1;
+    } else {
+      return null;
+    }
+  }
+  return parsed;
+}
+
+export type EvolveDocument = {
+  schema_version: 1;
+  product: string;
+  command: "evolve";
+} & (
+  | { state: "running" | "succeeded" | "failed"; run: Record<string, unknown> }
+  | { state: "rejected" | "unknown" | "unavailable"; code: string }
+);
+
+const RUN_ID = /^[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}$/;
+
+function evolveRun(value: unknown): Record<string, unknown> | null {
+  if (!value || typeof value !== "object") return null;
+  const run = value as Record<string, unknown>;
+  if (
+    typeof run.run_id !== "string" ||
+    !RUN_ID.test(run.run_id) ||
+    !["running", "succeeded", "failed"].includes(String(run.state))
+  ) {
+    return null;
+  }
+  return run;
+}
+
+export async function runEvolveCommand(
+  command: EvolveCommand,
+  options: {
+    userId?: string;
+    reader: CoreReader;
+    operations: CoreOperations;
+    sleep?: (ms: number) => Promise<void>;
+    pollMs?: number;
+  },
+): Promise<EvolveDocument> {
+  const base = {
+    schema_version: 1 as const,
+    product: PRODUCT,
+    command: "evolve" as const,
+  };
+  const body: Record<string, unknown> = {};
+  if (options.userId !== undefined) body.user_id = options.userId;
+  if (command.agentId !== undefined) body.agent_id = command.agentId;
+  if (command.maxBatches !== undefined) body.max_batches = command.maxBatches;
+  const started = await options.operations.post<unknown>(
+    "/api/v2/evolve",
+    body,
+  );
+  if (started.state !== "committed") {
+    return { ...base, state: started.state, code: started.code };
+  }
+  let run = evolveRun(started.data);
+  if (run === null)
+    return { ...base, state: "unknown", code: "response_invalid" };
+  const sleep =
+    options.sleep ??
+    ((ms: number) => new Promise((done) => setTimeout(done, ms)));
+  while (command.wait && run.state === "running") {
+    await sleep(options.pollMs ?? 2_000);
+    const read = await options.reader<unknown>(
+      `/api/v2/evolve/${run.run_id}`,
+      options.userId === undefined ? {} : { user_id: options.userId },
+    );
+    if (read.state !== "ok") {
+      return { ...base, state: "unavailable", code: read.code };
+    }
+    const next = evolveRun(read.data);
+    if (next === null)
+      return { ...base, state: "unknown", code: "response_invalid" };
+    run = next;
+  }
+  return {
+    ...base,
+    state: run.state as "running" | "succeeded" | "failed",
+    run: projectRun(run),
+  };
+}
+
+/** Counts and codes; SDK result bodies stay in the Core receipt. */
+function projectRun(run: Record<string, unknown>): Record<string, unknown> {
+  const keep = [
+    "run_id",
+    "user_id",
+    "agent_id",
+    "state",
+    "trigger",
+    "started_at",
+    "finished_at",
+    "rounds",
+    "tasks_processed",
+    "batch_failures",
+    "cycle_status",
+    "should_continue",
+    "continuation_reason",
+    "remaining_clusters",
+    "remaining_actionable_clusters",
+    "remaining_fresh_facts",
+    "capped",
+    "elapsed_ms",
+    "error_code",
+    "failure_kind",
+  ];
+  return Object.fromEntries(
+    keep.filter((key) => key in run).map((key) => [key, run[key]]),
+  );
+}
+
+export function evolveExitCode(document: EvolveDocument): number {
+  return document.state === "running" || document.state === "succeeded" ? 0 : 1;
+}
+
+export function renderEvolveText(document: EvolveDocument): string {
+  if ("code" in document) {
+    return document.code === "memory_mode_not_ultra"
+      ? "进化只在 ultra 模式下可用。切换：memhub settings mode ultra --yes\n"
+      : `进化未能开始：${document.code}\n`;
+  }
+  const run = document.run;
+  if (document.state === "running") {
+    return `进化已开始（${run.run_id}），在后台运行；加 --wait 等待结果。\n`;
+  }
+  if (document.state === "failed") {
+    return `进化失败（${run.run_id}）：${String(run.error_code ?? "unknown")}\n`;
+  }
+  return `进化完成（${run.run_id}）：${run.rounds} 轮，处理 ${run.tasks_processed} 项${
+    run.cycle_status ? `，周期 ${run.cycle_status}` : ""
+  }${run.capped ? "，已到轮数上限，可再次运行继续" : ""}\n`;
+}

memhub/apps/cli/src/memory-report.ts

@@ -14,6 +14,25 @@ import { type CoreReader, RESERVED_MEMORY_SEGMENTS } from "./core-client.js";
 import type { CoreOperations } from "./core-operations.js";
 import { PRODUCT } from "./version.js";
 
+/** A MemHub user id, the same shape Core accepts. `local` is the default. */
+const USER_ID = /^[A-Za-z0-9._-]{1,64}$/;
+export const DEFAULT_USER_ID = "local";
+
+export function isUserId(value: string): boolean {
+  return USER_ID.test(value);
+}
+
+/**
+ * The `user_id` to send, or nothing. The default user is never named on the
+ * wire, so a request without `--user-id` stays exactly what it was before
+ * multi-user support and still works against an older Core.
+ */
+export function userIdParameters(userId?: string): { user_id?: string } {
+  return userId === undefined || userId === DEFAULT_USER_ID
+    ? {}
+    : { user_id: userId };
+}
+
 /** Output budget: a long memory is shortened visibly, never dropped silently. */
 const CONTENT_BUDGET = 240;
 
@@ -101,10 +120,12 @@ export interface MemoryDetailDocument {
 export async function readMemoryPage(
   reader: CoreReader,
   query: MemoryQuery,
+  userId?: string,
 ): Promise<MemoryPageDocument> {
   const parameters: Record<string, string | number> = {
     offset: query.offset,
     limit: query.limit,
+    ...userIdParameters(userId),
   };
   if (query.status !== undefined) parameters.statuses = query.status;
   if (query.source !== undefined) parameters.source = query.source;
@@ -156,8 +177,12 @@ export async function readMemoryDetail(
   reader: CoreReader,
   memoryId: string,
   namespace: MemoryNamespace = {},
+  userId?: string,
 ): Promise<MemoryDetailDocument> {
-  const parameters = namespaceParameters(namespace);
+  const parameters = {
+    ...namespaceParameters(namespace),
+    ...userIdParameters(userId),
+  };
   const outcome = await reader<unknown>(
     `/api/v2/memories/${memoryId}`,
     Object.keys(parameters).length > 0 ? parameters : undefined,
@@ -236,6 +261,7 @@ export async function addMemory(
   operations: CoreOperations,
   text: string,
   requestId: string = randomUUID(),
+  userId?: string,
 ): Promise<MemoryAddDocument> {
   const base = {
     schema_version: 2 as const,
@@ -259,7 +285,10 @@ export async function addMemory(
     }
     throw error;
   }
-  const result = await operations.post<unknown>(request.path, request.body);
+  const result = await operations.post<unknown>(request.path, {
+    ...request.body,
+    ...userIdParameters(userId),
+  });
   if (result.state !== "committed") {
     return { ...base, state: result.state, code: result.code };
   }
@@ -273,6 +302,7 @@ export async function addMemory(
 export async function readMemoryAddStatus(
   reader: CoreReader,
   requestId: string,
+  userId?: string,
 ): Promise<MemoryAddDocument> {
   const base = {
     schema_version: 2 as const,
@@ -283,7 +313,11 @@ export async function readMemoryAddStatus(
   if (!UUID.test(requestId)) {
     return { ...base, state: "invalid", code: "invalid_request_id" };
   }
-  const outcome = await reader<unknown>(memoryAddStatusPath(requestId));
+  const user = userIdParameters(userId);
+  const outcome =
+    user.user_id === undefined
+      ? await reader<unknown>(memoryAddStatusPath(requestId))
+      : await reader<unknown>(memoryAddStatusPath(requestId), user);
   if (outcome.state === "not_found") {
     return { ...base, state: "not_found", code: "memory_add_not_found" };
   }

memhub/apps/cli/src/profile-versions-cli.test.ts

@@ -0,0 +1,166 @@
+import { beforeEach, describe, expect, it, vi } from "vitest";
+
+import { run } from "./cli.js";
+
+const transport = vi.hoisted(() => ({
+  get: vi.fn(),
+  post: vi.fn(),
+  put: vi.fn(),
+  patch: vi.fn(),
+  delete: vi.fn(),
+}));
+vi.mock("axios", () => ({
+  default: {
+    create: vi.fn(() => transport),
+    get: vi.fn(),
+    post: vi.fn(),
+    isAxiosError: () => false,
+  },
+}));
+
+const runtime = vi.hoisted(() => ({
+  inspectRuntimeStatus: vi.fn(),
+  discoverRuntimeAttachment: vi.fn(),
+}));
+vi.mock("@one-memory/runtime-control", async (importOriginal) => ({
+  ...(await importOriginal<typeof import("@one-memory/runtime-control")>()),
+  ...runtime,
+}));
+
+const VERSIONS = {
+  user_id: "wife",
+  request_id: "r-versions",
+  elapsed_ms: 2,
+  chains: [
+    {
+      memory_id: "l0-head",
+      content: "The user's name is Fei.",
+      status: "active",
+      memory_at: null,
+      gmt_created: 1_790_000_100,
+      basic_info: { name: "Fei" },
+      versions: [
+        {
+          memory_id: "l0-head",
+          content: "The user's name is Fei.",
+          memory_at: null,
+          gmt_created: 1_790_000_100,
+          current: true,
+        },
+        {
+          memory_id: "l0-old",
+          content: "The user's name is 费天翔.",
+          memory_at: null,
+          gmt_created: 1_790_000_000,
+          current: false,
+        },
+      ],
+    },
+  ],
+};
+
+beforeEach(() => {
+  vi.clearAllMocks();
+  for (const method of Object.values(transport)) method.mockReset();
+  runtime.inspectRuntimeStatus.mockResolvedValue({
+    state: "running",
+    code: "runtime_verified",
+    instanceId: "instance-1",
+    host: "127.0.0.1",
+    port: 8765,
+    version: "1.2.3",
+  });
+  runtime.discoverRuntimeAttachment.mockResolvedValue({
+    instanceId: "instance-1",
+    apiToken: "synthetic-runtime-token",
+    host: "127.0.0.1",
+    port: 8765,
+    servicePid: 4242,
+    health: {
+      status: "ok",
+      service: "one-memory-core",
+      version: "1.2.3",
+      instance_id: "instance-1",
+    },
+  });
+  transport.get.mockImplementation(async (url: string) =>
+    url === "/api/v2/user-profile/versions"
+      ? { status: 200, data: VERSIONS }
+      : { status: 404, data: {} },
+  );
+});
+
+describe("memhub profile --versions", () => {
+  it("reads the L0 version chain for the named user", async () => {
+    const result = await capturedRun([
+      "profile",
+      "--versions",
+      "--user-id",
+      "wife",
+      "--json",
+    ]);
+    expect(result.code).toBe(0);
+    expect(transport.get).toHaveBeenCalledWith(
+      "/api/v2/user-profile/versions",
+      { params: { user_id: "wife" } },
+    );
+    const document = JSON.parse(result.stdout);
+    expect(document.state).toBe("ok");
+    expect(document.chains[0].versions).toHaveLength(2);
+  });
+
+  it("marks the current version in text and names no user for local", async () => {
+    const result = await capturedRun(["profile", "--versions"]);
+    expect(result.code).toBe(0);
+    const [url, config] = transport.get.mock.calls[0] ?? [];
+    expect(url).toBe("/api/v2/user-profile/versions");
+    expect(config?.params?.user_id).toBeUndefined();
+    expect(result.stdout).toContain("* l0-head");
+    expect(result.stdout).toContain("  l0-old");
+    expect(result.stdout).toContain("费天翔");
+  });
+
+  it("refuses an unknown flag and a repeated --versions", async () => {
+    expect((await capturedRun(["profile", "--version-chain"])).code).toBe(2);
+    expect(
+      (await capturedRun(["profile", "--versions", "--versions"])).code,
+    ).toBe(2);
+    expect(transport.get).not.toHaveBeenCalled();
+  });
+
+  it("reports a malformed answer as invalid", async () => {
+    transport.get.mockResolvedValue({ status: 200, data: { chains: "x" } });
+    const result = await capturedRun(["profile", "--versions", "--json"]);
+    expect(result.code).toBe(1);
+    expect(JSON.parse(result.stdout).code).toBe("response_invalid");
+  });
+});
+
+async function capturedRun(arguments_: string[]) {
+  let stdout = "";
+  let stderr = "";
+  const stdoutSpy = vi
+    .spyOn(process.stdout, "write")
+    .mockImplementation((chunk) => {
+      stdout += String(chunk);
+      return true;
+    });
+  const stderrSpy = vi
+    .spyOn(process.stderr, "write")
+    .mockImplementation((chunk) => {
+      stderr += String(chunk);
+      return true;
+    });
+  try {
+    return {
+      code: await run(arguments_, {
+        ONE_MEMORY_DATA_DIR: "/synthetic/data",
+      }),
+      stdout,
+      stderr,
+    };
+  } finally {
+    stdoutSpy.mockRestore();
+    stderrSpy.mockRestore();
+  }
+}

memhub/apps/cli/src/source-progress-report.ts

@@ -14,6 +14,11 @@ import {
   validateProcessing,
   validateSources,
 } from "./core-responses.js";
+import {
+  describePipeline,
+  type PipelineSummary,
+  pipelineLine,
+} from "./status-report.js";
 import { PRODUCT } from "./version.js";
 
 type ReadState =
@@ -38,6 +43,21 @@ export interface SourcesDocument {
   state: ReadState["state"];
   code?: string;
   sources: Array<Record<string, unknown>>;
+  /** Every grant of every user (`GET /api/v1/source-grants`). Absent when
+   * Core is too old to list them; `sources` is the default user's view. */
+  grants?: SourceGrant[];
+}
+
+/** One source grant. Only Hermes has grants for users other than `local`,
+ * one per Hermes profile root. */
+export interface SourceGrant {
+  user_id: string;
+  source_id: string;
+  root_label: string | null;
+  root_path: string | null;
+  authorized: boolean;
+  revoked_at: string | null;
+  sync_enabled: boolean;
 }
 
 export interface SyncDocument {
@@ -69,6 +89,8 @@ export interface ProgressDocument {
     summary_state: ReadState;
     summary: Record<string, unknown> | null;
   };
+  /** The SDK mode and last evolve per user; absent on an older Core. */
+  pipeline?: PipelineSummary | null;
 }
 
 export async function readSources(
@@ -86,14 +108,66 @@ export async function readSources(
       sources: [],
     };
   }
+  const grants = await readSourceGrants(reader);
   return {
     schema_version: SCHEMA_VERSION,
     product: PRODUCT,
     state: "ok",
-    sources: result.data.map(projectSource),
+    // `/api/v1/sources` is the default user's view of every source.
+    sources: result.data.map((source) => ({
+      ...projectSource(source),
+      user_id: "local",
+    })),
+    ...(grants === null ? {} : { grants }),
   };
 }
 
+/** The grants of every user, or null when Core cannot list them (an older
+ * Core answers 404): the source list is then shown as before. */
+async function readSourceGrants(
+  reader: CoreReader,
+): Promise<SourceGrant[] | null> {
+  let outcome: ReadOutcome<unknown>;
+  try {
+    outcome = await reader<unknown>("/api/v1/source-grants");
+  } catch {
+    return null;
+  }
+  if (outcome.state !== "ok") return null;
+  const value = outcome.data;
+  if (!value || typeof value !== "object" || Array.isArray(value)) return null;
+  const items = (value as { grants?: unknown }).grants;
+  if (!Array.isArray(items)) return null;
+  const grants: SourceGrant[] = [];
+  const text = (candidate: unknown): string | null =>
+    typeof candidate === "string" && candidate.length <= 4096
+      ? candidate
+      : null;
+  for (const item of items) {
+    if (!item || typeof item !== "object" || Array.isArray(item)) return null;
+    const grant = item as Record<string, unknown>;
+    if (
+      typeof grant.user_id !== "string" ||
+      !/^[A-Za-z0-9._-]{1,64}$/.test(grant.user_id) ||
+      typeof grant.source_id !== "string" ||
+      !/^[a-z0-9][a-z0-9-]{0,127}$/.test(grant.source_id) ||
+      typeof grant.authorized !== "boolean"
+    ) {
+      return null;
+    }
+    grants.push({
+      user_id: grant.user_id,
+      source_id: grant.source_id,
+      root_label: text(grant.root_label),
+      root_path: text(grant.root_path),
+      authorized: grant.authorized,
+      revoked_at: text(grant.revoked_at),
+      sync_enabled: grant.sync_enabled === true,
+    });
+  }
+  return grants;
+}
+
 export async function readSync(reader: CoreReader): Promise<SyncDocument> {
   const result = validated(
     await reader<unknown>("/api/v1/live-sync"),
@@ -216,6 +290,10 @@ export async function readProgress(
       summary:
         processing.state === "ok" ? projectProcessing(processing.data) : null,
     },
+    pipeline:
+      processingRead.state === "ok"
+        ? describePipeline(processingRead.data)
+        : null,
   };
 }
 
@@ -231,6 +309,22 @@ export function renderSourcesText(document: SourcesDocument): string {
     );
   }
   if (document.sources.length === 0) lines.push("（没有来源）");
+  // The list above is the default user's (`local`). Another user's grant,
+  // a Hermes profile, is listed with its user so it is not mistaken for it.
+  const others = (document.grants ?? []).filter(
+    (grant) => grant.user_id !== "local",
+  );
+  if (others.length > 0) {
+    lines.push("其他用户的连接：");
+    for (const grant of others) {
+      const name =
+        document.sources.find((source) => source.source_id === grant.source_id)
+          ?.display_name ?? grant.source_id;
+      lines.push(
+        `- ${name}（用户 ${grant.user_id}）：${grant.authorized ? "已连接" : grant.revoked_at ? "已撤销" : "未连接"}；处理${grant.sync_enabled ? "已开启" : "已冻结"}${grant.root_path ? `；目录 ${grant.root_path}` : ""}`,
+      );
+    }
+  }
   lines.push("");
   return lines.join("\n");
 }
@@ -270,6 +364,18 @@ export function renderProgressText(document: ProgressDocument): string {
       ? `记忆整理：待处理 ${document.generation.summary.pending} · 进行中 ${document.generation.summary.running} · 等待模型 ${document.generation.summary.waiting_for_model} · 失败 ${document.generation.summary.failed}`
       : "记忆整理：暂时无法读取",
   );
+  if (document.pipeline) {
+    lines.push(`记忆模式：${pipelineLine(document.pipeline)}`);
+    for (const [userId, user] of Object.entries(
+      document.pipeline.evolve_users,
+    )) {
+      lines.push(
+        `  进化 ${userId}：${user.last_run_at ?? "尚未进化"}${
+          user.last_error ? `（失败：${user.last_error}）` : ""
+        }`,
+      );
+    }
+  }
   if (document.imports.output_truncated) {
     lines.push("这里只显示汇总；完整记录请使用 --json。");
   }

memhub/apps/cli/src/source-progress.test.ts

@@ -179,12 +179,72 @@ describe("CLI-07/10 source and synchronization reads", () => {
       sync_enabled: true,
       sync_completeness: "additions_only",
     });
-    expect(requestedPaths()).toEqual(["/api/v1/sources"]);
+    // The grants list answers 404 here, as an older Core does: the source
+    // list is shown as before, every row the default user's.
+    expect(report.sources[0].user_id).toBe("local");
+    expect(report.grants).toBeUndefined();
+    expect(requestedPaths()).toEqual([
+      "/api/v1/sources",
+      "/api/v1/source-grants",
+    ]);
     expect(requestedPaths().some((path) => path.includes("detect"))).toBe(
       false,
     );
   });
 
+  it("shows every user's grant, a Hermes profile with its user", async () => {
+    const grants = {
+      schema_version: 1,
+      grants: [
+        {
+          user_id: "local",
+          source_id: "codex",
+          root_label: ".codex",
+          root_path: "/home/synthetic/.codex",
+          authorized: true,
+          revoked_at: null,
+          sync_enabled: true,
+        },
+        {
+          user_id: "wife",
+          source_id: "hermes",
+          root_label: "wife",
+          root_path: "/home/synthetic/.hermes/profiles/wife",
+          authorized: true,
+          revoked_at: null,
+          sync_enabled: true,
+        },
+      ],
+    };
+    transport.get.mockImplementation(async (url: string) =>
+      url === "/api/v1/source-grants"
+        ? { status: 200, data: grants }
+        : (RESPONSES.get(url) ?? { status: 404, data: {} }),
+    );
+    let streams = captureStreams();
+    try {
+      expect(await run(["sources", "list", "--json"], {})).toBe(0);
+    } finally {
+      streams.restore();
+    }
+    const report = JSON.parse(streams.stdout());
+    expect(report.grants).toEqual(grants.grants);
+    expect(report.sources[0]).toMatchObject({
+      source_id: "codex",
+      user_id: "local",
+    });
+    streams = captureStreams();
+    try {
+      expect(await run(["sources", "list"], {})).toBe(0);
+    } finally {
+      streams.restore();
+    }
+    expect(streams.stdout()).toContain("其他用户的连接：");
+    expect(streams.stdout()).toContain(
+      "hermes（用户 wife）：已连接；处理已开启；目录 /home/synthetic/.hermes/profiles/wife",
+    );
+  });
+
   it("keeps degraded synchronization distinct and prints its error code", async () => {
     const streams = captureStreams();
     try {
@@ -552,6 +612,7 @@ describe("CLI-19 progress and issue reads", () => {
       requestedPaths().every((path) =>
         [
           "/api/v1/sources",
+          "/api/v1/source-grants",
           "/api/v1/live-sync",
           "/api/v1/import-jobs",
           "/api/v1/memory-processing",

memhub/apps/cli/src/status-report.ts

@@ -57,6 +57,9 @@ export interface StatusDocument {
     broker?: BackgroundHealth["broker"];
   };
   sections: Record<string, Section>;
+  /** R10: the SDK mode and System 2 state; null when the Runtime is not read
+   * or is older than the setting. */
+  pipeline?: PipelineSummary | null;
   summary: {
     state: "ok" | "partial";
     degraded: string[];
@@ -64,6 +67,99 @@ export interface StatusDocument {
   };
 }
 
+export interface PipelineSummary {
+  mode: "lite" | "pro" | "ultra";
+  configured_mode: "lite" | "pro" | "ultra";
+  forced_by_env: boolean;
+  last_evolve_at: string | null;
+  evolve_schedule: { enabled: boolean; time: string } | null;
+  evolve_running: number;
+  evolve_users: Record<
+    string,
+    {
+      last_run_at: string | null;
+      last_error: string | null;
+      state: string | null;
+    }
+  >;
+  graph_build: string | null;
+  /** SDK features (`memory_features`) whose value differs from the default. */
+  features: Record<string, boolean>;
+}
+
+const MODES = new Set(["lite", "pro", "ultra"]);
+
+/** Projection of `pipeline` in `/api/v1/memory-processing`; null if absent or malformed. */
+export function describePipeline(value: unknown): PipelineSummary | null {
+  if (!value || typeof value !== "object") return null;
+  const pipeline = (value as { pipeline?: unknown }).pipeline;
+  if (!pipeline || typeof pipeline !== "object") return null;
+  const item = pipeline as Record<string, unknown>;
+  if (
+    !MODES.has(String(item.mode)) ||
+    !MODES.has(String(item.configured_mode))
+  ) {
+    return null;
+  }
+  const evolve =
+    item.evolve && typeof item.evolve === "object"
+      ? (item.evolve as Record<string, unknown>)
+      : {};
+  const schedule =
+    evolve.schedule && typeof evolve.schedule === "object"
+      ? (evolve.schedule as Record<string, unknown>)
+      : null;
+  const users: PipelineSummary["evolve_users"] = {};
+  if (evolve.users && typeof evolve.users === "object") {
+    for (const [userId, raw] of Object.entries(
+      evolve.users as Record<string, unknown>,
+    )) {
+      const user =
+        raw && typeof raw === "object" ? (raw as Record<string, unknown>) : {};
+      const result =
+        user.last_result && typeof user.last_result === "object"
+          ? (user.last_result as Record<string, unknown>)
+          : {};
+      users[userId] = {
+        last_run_at:
+          typeof user.last_run_at === "string" ? user.last_run_at : null,
+        last_error:
+          typeof user.last_error === "string" ? user.last_error : null,
+        state: typeof result.state === "string" ? result.state : null,
+      };
+    }
+  }
+  const features: Record<string, boolean> = {};
+  if (item.features && typeof item.features === "object") {
+    for (const [name, value] of Object.entries(
+      item.features as Record<string, unknown>,
+    )) {
+      if (typeof value === "boolean") features[name] = value;
+    }
+  }
+  const graph =
+    item.graph_build && typeof item.graph_build === "object"
+      ? (item.graph_build as Record<string, unknown>)
+      : {};
+  return {
+    mode: item.mode as PipelineSummary["mode"],
+    configured_mode: item.configured_mode as PipelineSummary["mode"],
+    forced_by_env: item.forced_by_env === true,
+    last_evolve_at:
+      typeof item.last_evolve_at === "string" ? item.last_evolve_at : null,
+    evolve_schedule:
+      schedule &&
+      typeof schedule.enabled === "boolean" &&
+      typeof schedule.time === "string"
+        ? { enabled: schedule.enabled, time: schedule.time }
+        : null,
+    evolve_running: Array.isArray(evolve.running) ? evolve.running.length : 0,
+    evolve_users: users,
+    graph_build: typeof graph.state === "string" ? graph.state : null,
+    features,
+  };
+}
+
 const SECTION_NAMES = [
   "sources",
   "sync",
@@ -97,6 +193,7 @@ export async function composeStatusReport(options: {
   attachFailure?: string;
 }): Promise<StatusDocument> {
   const sections: Record<string, Section> = {};
+  let pipeline: PipelineSummary | null = null;
   if (options.reader === null) {
     const state: SectionState =
       options.attachFailure === undefined ? "not_checked" : "unavailable";
@@ -118,6 +215,8 @@ export async function composeStatusReport(options: {
       validateProcessing,
       describeProcessing,
     );
+    pipeline =
+      processing.state === "ok" ? describePipeline(processing.data) : null;
     sections.model = section(model, validateModelProfiles, describeModel);
     sections.memories = section(
       memories,
@@ -169,6 +268,7 @@ export async function composeStatusReport(options: {
     },
     bridge,
     sections,
+    pipeline,
     summary: {
       state:
         degraded.length === 0 && notChecked.length === 0 ? "ok" : "partial",
@@ -205,6 +305,7 @@ export function renderStatusText(
       `${SECTION_LABELS[name]}：${sectionLine(report.sections[name])}`,
     );
   }
+  if (report.pipeline) lines.push(`记忆模式：${pipelineLine(report.pipeline)}`);
   if (report.summary.state === "partial") {
     const notes = [
       report.summary.degraded.length > 0
@@ -220,6 +321,36 @@ export function renderStatusText(
   return lines.join("\n");
 }
 
+export function pipelineLine(pipeline: PipelineSummary): string {
+  const parts: string[] = [pipeline.mode];
+  if (pipeline.forced_by_env)
+    parts.push("（ONE_MEMORY_GRAPH_STORE 强制 ultra）");
+  if (pipeline.mode === "ultra") {
+    const schedule = pipeline.evolve_schedule;
+    parts.push(
+      schedule?.enabled
+        ? `，每日 ${schedule.time} 自动进化`
+        : "，自动进化已关闭",
+    );
+    parts.push(
+      pipeline.last_evolve_at
+        ? `，上次进化 ${pipeline.last_evolve_at}`
+        : "，尚未进化",
+    );
+    if (pipeline.evolve_running > 0)
+      parts.push(`，进行中 ${pipeline.evolve_running}`);
+  }
+  const changed = Object.entries(pipeline.features);
+  if (changed.length > 0) {
+    parts.push(
+      `；SDK 功能：${changed
+        .map(([name, value]) => `${name}=${value ? "开" : "关"}`)
+        .join("、")}`,
+    );
+  }
+  return parts.join("");
+}
+
 function runtimeLine(report: StatusDocument): string {
   if (report.runtime.state !== "running") {
     return report.runtime.state === "stopped" ? "未运行" : "暂时无法确认";

memhub/apps/cli/src/user-id-cli.test.ts

@@ -0,0 +1,405 @@
+import { mkdtemp, writeFile } from "node:fs/promises";
+import os, { tmpdir } from "node:os";
+import path from "node:path";
+import { beforeEach, describe, expect, it, vi } from "vitest";
+
+import { run } from "./cli.js";
+
+const transport = vi.hoisted(() => ({
+  get: vi.fn(),
+  post: vi.fn(),
+  put: vi.fn(),
+  patch: vi.fn(),
+  delete: vi.fn(),
+}));
+const direct = vi.hoisted(() => ({ get: vi.fn(), post: vi.fn() }));
+vi.mock("axios", () => ({
+  default: {
+    create: vi.fn(() => transport),
+    get: direct.get,
+    post: direct.post,
+    isAxiosError: () => false,
+  },
+}));
+
+const runtime = vi.hoisted(() => ({
+  inspectRuntimeStatus: vi.fn(),
+  discoverRuntimeAttachment: vi.fn(),
+}));
+vi.mock("@one-memory/runtime-control", async (importOriginal) => ({
+  ...(await importOriginal<typeof import("@one-memory/runtime-control")>()),
+  ...runtime,
+}));
+
+const REQUEST = "33333333-3333-4333-8333-333333333333";
+const HERMES = {
+  source_id: "hermes",
+  display_name: "Hermes",
+  storage_kinds: ["sqlite"],
+  available: true,
+  authorized: true,
+  detected: null,
+  root_label: ".hermes",
+  candidate_path: "~/.hermes",
+  candidate_kind: "default",
+  candidate_revision: 0,
+  candidate_root_id: "hermes-root-identity",
+  imported_messages: 0,
+  imported_conversations: 0,
+  capabilities: {
+    historical_import: "implemented",
+    incremental_import: "implemented",
+    live_sync: "implemented",
+    recall: "implemented",
+    installer: "not_implemented",
+    verification: "fixture_only",
+  },
+  sync_completeness: "additions_only",
+  location_state: "found",
+};
+const CODEX = {
+  ...HERMES,
+  source_id: "codex",
+  display_name: "Codex",
+  candidate_root_id: "codex-root-identity",
+};
+
+function addReceipt() {
+  return {
+    evidence_id: "44444444-4444-4444-8444-444444444444",
+    text: "Prefers tea",
+    scope: "all-apps",
+    created_at: "2026-10-27T00:00:00Z",
+    state: "pending",
+    task_id: null,
+    memory_ids: [],
+    error: null,
+  };
+}
+
+beforeEach(() => {
+  vi.clearAllMocks();
+  for (const method of [
+    ...Object.values(transport),
+    ...Object.values(direct),
+  ]) {
+    method.mockReset();
+  }
+  runtime.inspectRuntimeStatus.mockResolvedValue({
+    state: "running",
+    code: "runtime_verified",
+    instanceId: "instance-1",
+    host: "127.0.0.1",
+    port: 8765,
+    version: "1.2.3",
+  });
+  runtime.discoverRuntimeAttachment.mockResolvedValue({
+    instanceId: "instance-1",
+    apiToken: "synthetic-runtime-token",
+    host: "127.0.0.1",
+    port: 8765,
+    servicePid: 4242,
+    health: {
+      status: "ok",
+      service: "one-memory-core",
+      version: "1.2.3",
+      instance_id: "instance-1",
+    },
+  });
+  transport.get.mockImplementation(async (url: string) => {
+    if (url === "/api/v1/sources") {
+      return { status: 200, data: [HERMES, CODEX] };
+    }
+    if (url === "/api/v2/user-profile") {
+      return {
+        status: 200,
+        data: { request_id: "r-1", profile: null, elapsed_ms: 1 },
+      };
+    }
+    return { status: 404, data: {} };
+  });
+  transport.post.mockImplementation(
+    async (url: string, body: Record<string, unknown>) =>
+      url === "/api/v1/sources/connect"
+        ? {
+            status: 200,
+            data: {
+              request_id: body.request_id,
+              accepted_source_ids: ["hermes"],
+              failed_sources: {},
+              jobs: [],
+            },
+          }
+        : url === "/api/v2/memories"
+          ? { status: 201, data: addReceipt() }
+          : { status: 404, data: {} },
+  );
+  transport.put.mockResolvedValue({ status: 200, data: {} });
+  transport.delete.mockResolvedValue({ status: 204, data: "" });
+  direct.post.mockResolvedValue({
+    status: 202,
+    data: {
+      request_id: REQUEST,
+      state: "accepted",
+      created_at: "T1",
+      updated_at: "T1",
+    },
+  });
+  direct.get.mockResolvedValue({
+    data: {
+      request_id: REQUEST,
+      state: "accepted",
+      created_at: "T1",
+      updated_at: "T1",
+    },
+  });
+});
+
+describe("--user-id on the memory commands", () => {
+  it("passes the user as a query parameter to memory reads and deletes", async () => {
+    await capturedRun(["memories", "list", "--user-id", "wife", "--json"]);
+    expect(transport.get).toHaveBeenCalledWith("/api/v2/memory-library", {
+      params: { offset: 0, limit: 100, user_id: "wife" },
+    });
+    // Global: the flag may come before the subcommand too.
+    await capturedRun(["memories", "--user-id", "wife", "show", "m-1"]);
+    expect(transport.get).toHaveBeenCalledWith("/api/v2/memories/m-1", {
+      params: { user_id: "wife" },
+    });
+    await capturedRun(["memories", "add-status", REQUEST, "--user-id", "wife"]);
+    expect(transport.get).toHaveBeenCalledWith(
+      `/api/v2/memories/additions/${REQUEST}`,
+      { params: { user_id: "wife" } },
+    );
+    await capturedRun(["profile", "--user-id", "wife", "--json"]);
+    expect(transport.get).toHaveBeenCalledWith("/api/v2/user-profile", {
+      params: { user_id: "wife" },
+    });
+    await capturedRun([
+      "memories",
+      "delete",
+      "fact-1",
+      "--kind",
+      "fact",
+      "--yes",
+      "--user-id",
+      "wife",
+    ]);
+    expect(transport.delete).toHaveBeenCalledWith(
+      "/api/v2/memories/fact-1?kind=fact&user_id=wife",
+      undefined,
+    );
+    await capturedRun([
+      "memories",
+      "delete",
+      "m-1",
+      "--yes",
+      "--user-id",
+      "wife",
+    ]);
+    expect(transport.delete).toHaveBeenCalledWith(
+      "/api/v2/memories/m-1?user_id=wife",
+      undefined,
+    );
+    await capturedRun(["write", "status", REQUEST, "--user-id", "wife"]);
+    expect(direct.get.mock.calls[0][0]).toBe(
+      `http://127.0.0.1:8765/api/v2/memory-writes/${REQUEST}`,
+    );
+    expect(direct.get.mock.calls[0][1].params).toEqual({ user_id: "wife" });
+  });
+
+  it("passes the user in the body of memory writes and searches", async () => {
+    await capturedRun(["memories", "add", "Prefers tea", "--user-id", "wife"]);
+    expect(transport.post.mock.calls[0][0]).toBe("/api/v2/memories");
+    expect(transport.post.mock.calls[0][1]).toMatchObject({
+      text: "Prefers tea",
+      scope: "all-apps",
+      user_id: "wife",
+    });
+    const directory = await mkdtemp(path.join(tmpdir(), "om-user-id-"));
+    const content = path.join(directory, "content.txt");
+    await writeFile(content, "Prefers green tea\n");
+    await capturedRun([
+      "memories",
+      "update",
+      "m-1",
+      "--content-file",
+      content,
+      "--user-id",
+      "wife",
+    ]);
+    expect(transport.put.mock.calls[0][0]).toBe("/api/v2/memories/m-1");
+    expect(transport.put.mock.calls[0][1]).toMatchObject({ user_id: "wife" });
+    await capturedRun(["recall", "tea", "--user-id", "wife", "--json"]);
+    const search = transport.post.mock.calls.find(
+      ([url]) => url === "/api/v2/context/search",
+    );
+    expect(search?.[1]).toMatchObject({ query: "tea", user_id: "wife" });
+    const messages = path.join(directory, "messages.json");
+    await writeFile(
+      messages,
+      JSON.stringify([{ role: "user", content: "I like tea" }]),
+    );
+    await capturedRun([
+      "write",
+      "submit",
+      "--messages-file",
+      messages,
+      "--session",
+      "s-1",
+      "--user-id",
+      "wife",
+    ]);
+    expect(direct.post.mock.calls[0][0]).toBe(
+      "http://127.0.0.1:8765/api/v2/memory-writes",
+    );
+    expect(direct.post.mock.calls[0][1]).toMatchObject({
+      session_id: "s-1",
+      user_id: "wife",
+    });
+  });
+
+  it("names no user by default or for --user-id local", async () => {
+    for (const extra of [[], ["--user-id", "local"]]) {
+      transport.get.mockClear();
+      transport.post.mockClear();
+      await capturedRun(["memories", "list", ...extra, "--json"]);
+      expect(transport.get).toHaveBeenCalledWith("/api/v2/memory-library", {
+        params: { offset: 0, limit: 100 },
+      });
+      await capturedRun(["profile", ...extra]);
+      expect(transport.get).toHaveBeenCalledWith(
+        "/api/v2/user-profile",
+        undefined,
+      );
+      await capturedRun(["memories", "add", "Prefers tea", ...extra]);
+      expect(transport.post.mock.calls[0][1]).not.toHaveProperty("user_id");
+      await capturedRun(["recall", "tea", ...extra, "--json"]);
+      const search = transport.post.mock.calls.find(
+        ([url]) => url === "/api/v2/context/search",
+      );
+      expect(search?.[1]).not.toHaveProperty("user_id");
+    }
+  });
+
+  it("refuses a malformed user id before contacting Core", async () => {
+    for (const value of ["bad:id", "", "x".repeat(65), "a b", "--json"]) {
+      const result = await capturedRun([
+        "memories",
+        "list",
+        "--user-id",
+        value,
+      ]);
+      expect(result.code).toBe(2);
+    }
+    expect(
+      (await capturedRun(["profile", "--user-id", "a", "--user-id", "b"])).code,
+    ).toBe(2);
+    expect((await capturedRun(["recall", "tea", "--user-id"])).code).toBe(2);
+    expect(transport.get).not.toHaveBeenCalled();
+    expect(transport.post).not.toHaveBeenCalled();
+  });
+});
+
+describe("sources connect for a Hermes profile", () => {
+  it("connects hermes for another user from its own root", async () => {
+    const result = await capturedRun([
+      "sources",
+      "connect",
+      "hermes",
+      "--user-id",
+      "wife",
+      "--root",
+      "~/.hermes/profiles/wife",
+      "--json",
+    ]);
+    expect(result.code).toBe(0);
+    expect(transport.post).toHaveBeenCalledTimes(1);
+    const [url, body] = transport.post.mock.calls[0];
+    expect(url).toBe("/api/v1/sources/connect");
+    expect(body).toMatchObject({
+      source_ids: ["hermes"],
+      // Core reads the custom root's identity itself.
+      source_roots: {},
+      source_candidate_revisions: {},
+      onboarding: false,
+      user_id: "wife",
+      root: path.join(os.homedir(), ".hermes/profiles/wife"),
+    });
+  });
+
+  it("keeps a plain connect unchanged and accepts --import with a profile", async () => {
+    await capturedRun(["sources", "connect", "hermes", "--import", "--json"]);
+    const plain = transport.post.mock.calls[0][1];
+    expect(plain).not.toHaveProperty("user_id");
+    expect(plain).not.toHaveProperty("root");
+    expect(plain).toMatchObject({
+      source_roots: { hermes: "hermes-root-identity" },
+      onboarding: true,
+    });
+    transport.post.mockClear();
+    const result = await capturedRun([
+      "sources",
+      "connect",
+      "hermes",
+      "--import",
+      "--user-id",
+      "wife",
+      "--root",
+      "/profiles/wife",
+      "--request-id",
+      REQUEST,
+      "--json",
+    ]);
+    expect(result.code).toBe(0);
+    expect(transport.post.mock.calls[0][1]).toMatchObject({
+      request_id: REQUEST,
+      user_id: "wife",
+      root: "/profiles/wife",
+    });
+  });
+
+  it("refuses --user-id and --root for any source but hermes", async () => {
+    for (const argv of [
+      ["sources", "connect", "codex", "--user-id", "wife"],
+      ["sources", "connect", "codex", "--root", "/somewhere", "--import"],
+      ["sources", "connect", "hermes", "codex", "--user-id", "wife"],
+      ["sources", "connect", "--all-detected", "--user-id", "wife"],
+      ["sources", "connect", "hermes", "--user-id", "bad:id"],
+    ]) {
+      const result = await capturedRun(argv);
+      expect(result.code).toBe(2);
+      expect(result.stderr).toContain("hermes");
+    }
+    expect(transport.post).not.toHaveBeenCalled();
+  });
+});
+
+async function capturedRun(arguments_: string[]) {
+  let stdout = "";
+  let stderr = "";
+  const stdoutSpy = vi
+    .spyOn(process.stdout, "write")
+    .mockImplementation((chunk) => {
+      stdout += String(chunk);
+      return true;
+    });
+  const stderrSpy = vi
+    .spyOn(process.stderr, "write")
+    .mockImplementation((chunk) => {
+      stderr += String(chunk);
+      return true;
+    });
+  try {
+    return {
+      code: await run(arguments_, {
+        ONE_MEMORY_DATA_DIR: "/synthetic/data",
+      }),
+      stdout,
+      stderr,
+    };
+  } finally {
+    stdoutSpy.mockRestore();
+    stderrSpy.mockRestore();
+  }
+}

memhub/apps/core/pyproject.toml

@@ -12,7 +12,7 @@ dependencies = [
     "fastapi>=0.141.1",
     "filelock>=3.32.4",
     # Public SDK version validated against MemHub's permission and recall contracts.
-    "hy-memory==2.0.6",
+    "hy-memory==2.0.9",
     # chromadb arrives through the HY SDK; locked here so an unrelated upgrade
     # cannot move the SDK's store. Built-in embedding consumes numpy directly.
     "chromadb==1.5.9",

memhub/apps/core/src/one_memory/cli.py

@@ -584,6 +584,7 @@ def _sdk_migration_status(options: Sequence[str], stream: TextIO) -> int:
         }
     else:
         from one_memory.hy_runtime_config import HY_MEMORY_NAMESPACE
+        from one_memory.hy_sdk_runtime import builtin_retrieval_snapshot
         from one_memory.sdk_data_plane_migration import (
             MemoryAuthority,
             SdkDataPlaneMigrationStore,
@@ -595,14 +596,14 @@ def _sdk_migration_status(options: Sequence[str], stream: TextIO) -> int:
 
         authority = SdkDataPlaneMigrationStore(database_path).authority()
         if authority is MemoryAuthority.SDK:
-            profile = read_active_profile(database_path)
+            profile = read_active_profile(database_path) or builtin_retrieval_snapshot()
             leftovers = bool(
                 legacy_tables_present(database_path)
                 or abandoned_staging_present(
                     data_dir=settings.data_dir,
                     namespace=HY_MEMORY_NAMESPACE,
                 )
-                or (profile is not None and legacy_vector_collections(settings.data_dir, profile))
+                or legacy_vector_collections(settings.data_dir, profile)
             )
             state = "required" if leftovers else "not_required"
         else:

memhub/apps/core/src/one_memory/context_models.py

@@ -5,6 +5,9 @@ from typing import Annotated, Literal
 
 from pydantic import BaseModel, ConfigDict, Field, model_validator
 
+from one_memory.sdk_search_options import SdkSearchOptions
+from one_memory.user_scope import DEFAULT_USER_ID, UserId
+
 ContextKind = Literal["raw", "memory"]
 ContextSearchChannel = Literal["lexical", "semantic"]
 ContextSemanticStatus = Literal["not_searched", "completed", "failed", "timed_out"]
@@ -47,17 +50,24 @@ class ContextSearchRequest(BaseModel):
         return self
 
 
-class SdkContextSearchRequest(ContextSearchRequest):
+class SdkContextSearchRequest(ContextSearchRequest, SdkSearchOptions):
     """`POST /api/v2/context/search`: the three SDK recall lanes are set apart.
 
     `limit` is the normal lane's top-k and the cap on returned items;
     `schema_limit` and `intention_limit` are the SDK's schema (L6) and
     proactive intention (L7) lanes. A caller's values are passed through.
+
+    The `SdkSearchOptions` fields reach the memory lane's `client.search`;
+    the raw lane takes the floor and the source, session, agent and time
+    filters (never wider than the caller's raw authority). Absent, both
+    lanes make the call they always made.
     """
 
     limit: int = Field(default=10, ge=1, le=20, strict=True)
     schema_limit: int = Field(default=3, ge=0, le=20, strict=True)
     intention_limit: int = Field(default=1, ge=0, le=20, strict=True)
+    #: Whose memories to search. The App never sends it and reads `local`.
+    user_id: UserId = DEFAULT_USER_ID
 
 
 class ContextMessage(BaseModel):

memhub/apps/core/src/one_memory/context_routes.py

@@ -24,13 +24,57 @@ from one_memory.hy_runtime_config import HY_MEMORY_READER
 from one_memory.hy_sdk_runtime import sdk_relevance_floor
 from one_memory.hy_sdk_types import (
     HySdkMemoryRecord,
+    parse_sdk_lane_search,
     parse_sdk_raw_search,
     parse_sdk_search,
 )
 from one_memory.imports.storage import EvidenceStore
+from one_memory.sdk_search_options import SdkSearchFilters, SearchLane
+from one_memory.semantic_deadline import SemanticMemoryHit
+from one_memory.user_scope import owns
 
 CALL_EVENT_RESPONSE_BUDGET_MS = 250
 
+#: Pass-through filters the raw lane honours. Scene, layers, reader and the
+#: schema/working floors describe memory lanes; the raw lane is always the
+#: chat store's L1 raw layer.
+_RAW_LANE_FILTERS = ("created_after", "session_ids", "agent_ids", "exclude_sources", "rerank")
+
+
+def raw_lane_sources(
+    allowed: Sequence[str] | frozenset[str] | None,
+    requested: Sequence[str] | None,
+) -> list[str] | None:
+    """The raw lane's source filter: the caller's, never wider than its authority.
+
+    `allowed` empty or None means every source (the call Core always made).
+    None back means nothing is left to search.
+    """
+    allowed_list = list(allowed or ())
+    if requested is None:
+        return allowed_list
+    chosen = [
+        source for source in dict.fromkeys(requested) if not allowed_list or source in allowed_list
+    ]
+    return chosen or None
+
+
+def raw_lane_overrides(value: SdkSearchFilters) -> dict[str, object]:
+    """The caller's raw-lane overrides: its floor and its row filters."""
+    overrides = value.filter_kwargs()
+    found = {name: overrides[name] for name in _RAW_LANE_FILTERS if name in overrides}
+    floor = value.lane_floor("raw")
+    if floor is not None:
+        found["min_score"] = floor
+    return found
+
+
+def hit_namespace(hit: SemanticMemoryHit) -> dict[str, str]:
+    """The `get` namespace of a hit: nothing for chat, kind and variant otherwise."""
+    if hit.kind in {"experience", "fact"} and hit.variant is not None:
+        return {"kind": hit.kind, "variant": hit.variant}
+    return {}
+
 
 def sdk_context_router(dependencies: Sequence[Depends]) -> APIRouter:
     """SDK-native context: no Evidence/Governance ids or fabricated revisions."""
@@ -66,26 +110,36 @@ def sdk_context_router(dependencies: Sequence[Depends]) -> APIRouter:
             client = await client_handle.client()
         except Exception:
             raise HTTPException(503, "sdk_context_unavailable") from None
+        memory_mode = getattr(request.app.state, "memory_mode", None)
+        lite = callable(memory_mode) and memory_mode() == "lite"
 
         async def raw_lane() -> tuple[SdkRawContextItem, ...]:
             if "raw" not in value.kinds or not authority.raw_allowed:
                 return ()
+            sources = raw_lane_sources(authority.raw_sources, value.include_sources)
+            if sources is None:
+                return ()
             raw = await asyncio.to_thread(
                 client.search,
                 value.query,
-                user_ids=["local"],
-                agent_ids=[],
-                include_sources=list(authority.raw_sources or ()),
-                limit=value.limit,
-                min_score=0.0,
-                schema_min_score=0.0,
-                schema_limit=0,
-                intention_limit=0,
-                scene="chat",
-                reader=HY_MEMORY_READER,
-                layers=["l1_raw"],
+                **{
+                    "user_ids": [value.user_id],
+                    "agent_ids": [],
+                    "include_sources": sources,
+                    "limit": value.limit,
+                    "min_score": 0.0,
+                    "schema_min_score": 0.0,
+                    "schema_limit": 0,
+                    "intention_limit": 0,
+                    "scene": "chat",
+                    "reader": HY_MEMORY_READER,
+                    "layers": ["l1_raw"],
+                    **raw_lane_overrides(value),
+                },
+            )
+            hits = tuple(
+                hit for hit in parse_sdk_raw_search(raw) if value.passes_floor("raw", hit.score)
             )
-            hits = parse_sdk_raw_search(raw)
             return tuple(
                 SdkRawContextItem(
                     raw_id=hit.raw_id,
@@ -103,27 +157,48 @@ def sdk_context_router(dependencies: Sequence[Depends]) -> APIRouter:
             )
 
         async def memory_lane() -> tuple[SdkMemoryContextItem, ...]:
-            if "memory" not in value.kinds or not authority.memory_allowed:
+            # lite extracts nothing: its search answers raw turns, which the
+            # raw lane already reads under the raw authority.
+            if "memory" not in value.kinds or not authority.memory_allowed or lite:
                 return ()
+            overrides = value.search_overrides()
             raw = await asyncio.to_thread(
                 client.search,
                 value.query,
-                user_ids=["local"],
-                agent_ids=[],
-                limit=value.limit,
-                min_score=sdk_relevance_floor(),
-                schema_min_score=sdk_relevance_floor(),
-                schema_limit=value.schema_limit,
-                intention_limit=value.intention_limit,
-                scene="normal",
-                reader=HY_MEMORY_READER,
+                **{
+                    "user_ids": [value.user_id],
+                    "agent_ids": [],
+                    "limit": value.limit,
+                    "min_score": sdk_relevance_floor(),
+                    "schema_min_score": sdk_relevance_floor(),
+                    "schema_limit": value.schema_limit,
+                    "intention_limit": value.intention_limit,
+                    "scene": "normal",
+                    "reader": HY_MEMORY_READER,
+                    **overrides,
+                },
+            )
+            hits = (
+                tuple(
+                    hit
+                    for hit in parse_sdk_lane_search(
+                        raw, error_code="hy_memory_search_shape_unknown"
+                    )
+                    if value.passes_floor(cast(SearchLane, hit.lane), hit.score)
+                )
+                if overrides
+                else parse_sdk_search(raw, error_code="hy_memory_search_shape_unknown")
             )
-            hits = parse_sdk_search(raw, error_code="hy_memory_search_shape_unknown")
             items: list[SdkMemoryContextItem] = []
             for hit in hits:
-                record = await asyncio.to_thread(client.get, hit.memory_id)
+                record = await asyncio.to_thread(client.get, hit.memory_id, **hit_namespace(hit))
                 memory = None if record is None else HySdkMemoryRecord.from_sdk(record)
-                if memory is None or memory.status != "active":
+                # `get` takes no user: a hit is kept only if it is the caller's.
+                if (
+                    memory is None
+                    or memory.status != "active"
+                    or not owns(memory.user_id, value.user_id)
+                ):
                     continue
                 items.append(
                     SdkMemoryContextItem(

memhub/apps/core/src/one_memory/hy_graph_admin.py

@@ -71,6 +71,16 @@ def graph_store_enabled() -> bool:
     return raw != "none"
 
 
+def effective_memory_mode(configured: str = "pro") -> str:
+    """The SDK mode clients are built with: the setting, unless dev forces ultra.
+
+    `ONE_MEMORY_GRAPH_STORE` (any value but empty or `none`) predates the
+    setting and keeps working for development: it forces ultra whatever the
+    configured mode. Without it the configured mode (default pro) applies.
+    """
+    return "ultra" if graph_store_enabled() else configured
+
+
 def inspect_graph_width_mismatch(config: Any) -> GraphWidthMismatch | None:
     """Return a mismatch without renaming or deleting anything."""
     db_path = Path(config.graph_store.db_path)
@@ -83,11 +93,12 @@ def inspect_profile_graph_width(
     *,
     api_key: str | None,
     data_dir: Path,
+    mode: str | None = None,
 ) -> GraphWidthMismatch | None:
     """Inspect the graph path and width selected by the active profile."""
     del api_key
     return _inspect_graph_width(
-        _profile_graph_path(data_dir),
+        _profile_graph_path(data_dir, mode),
         _profile_embedding_width(profile),
     )
 
@@ -98,11 +109,12 @@ def decide_profile_graph_width(
     api_key: str | None,
     data_dir: Path,
     decision: GraphWidthDecision,
+    mode: str | None = None,
 ) -> GraphWidthOutcome:
     """Apply the user's decision to the active profile's graph."""
     del api_key
     config = _GraphAdminConfig(
-        graph_store=_GraphStoreConfig(str(_profile_graph_path(data_dir))),
+        graph_store=_GraphStoreConfig(str(_profile_graph_path(data_dir, mode))),
         embedder=_EmbeddingWidthConfig(_profile_embedding_width(profile)),
     )
     return apply_graph_width_decision(
@@ -240,8 +252,9 @@ def _inspect_graph_width(
     )
 
 
-def _profile_graph_path(data_dir: Path) -> Path:
-    name = "graph" if graph_store_enabled() else "graph-disabled"
+def _profile_graph_path(data_dir: Path, mode: str | None = None) -> Path:
+    ultra = mode == "ultra" if mode is not None else graph_store_enabled()
+    name = "graph" if ultra else "graph-disabled"
     return data_dir / "hy-memory" / HY_MEMORY_NAMESPACE / name
 
 

memhub/apps/core/src/one_memory/hy_runtime_config.py

@@ -14,7 +14,7 @@ from pathlib import Path
 from typing import Literal, cast
 
 HY_MEMORY_DISTRIBUTION = "hy-memory"
-HY_MEMORY_VERSION = "2.0.6"
+HY_MEMORY_VERSION = "2.0.9"
 HY_MEMORY_NAMESPACE = "governed-v2"
 DiagnosticLogLevel = Literal["error", "warning", "info", "debug"]
 
@@ -109,11 +109,30 @@ def apply_diagnostic_log_level(level: DiagnosticLogLevel) -> None:
         if any(name == root or name.startswith(f"{root}.") for root in _NOISY_LIBRARY_LOGGERS):
             logging.getLogger(name).setLevel(noisy_level)
 
+    # HY configures its text log once, on the first client, at the level of
+    # that moment. Its file handler must follow later changes too, or a live
+    # switch to debug never reaches hy_memory.log.
+    for handler in logging.getLogger("hy_memory").handlers:
+        if isinstance(handler, logging.FileHandler):
+            handler.setLevel(selected)
+
     os.environ["MEMORY_LOG_LEVEL"] = logging.getLevelName(selected)
-    os.environ["MEMORY_PIPELINE_TRACE_ENABLED"] = "true" if level == "debug" else "false"
+    _apply_hy_pipeline_policy()
     os.environ["MEMORY_LOG_LLM_CONTENT"] = "true" if level == "debug" else "false"
 
 
+def _apply_hy_pipeline_policy() -> None:
+    """Keep HY's per-step pipeline log and trace off at every level.
+
+    Nothing in MemHub reads them: Core's own diagnostics carry the step
+    timing, and the JSONL/SQLite copies only grew on disk. With both off
+    the SDK skips the per-step bookkeeping altogether.
+    """
+
+    os.environ["MEMORY_PIPELINE_LOG_ENABLED"] = "false"
+    os.environ["MEMORY_PIPELINE_TRACE_ENABLED"] = "false"
+
+
 def hy_memory_version() -> str:
     """Return the installed SDK version, failing if the pinned dist is absent."""
 
@@ -131,7 +150,9 @@ def configure_hy_runtime_environment(data_dir: Path) -> None:
     os.environ["MEMORY_LOG_DIR"] = str(log_dir)
     level = diagnostic_log_level(data_dir)
     os.environ["MEMORY_LOG_LEVEL"] = level.upper()
-    os.environ["MEMORY_PIPELINE_TRACE_ENABLED"] = "true" if level == "debug" else "false"
+    _apply_hy_pipeline_policy()
+    # Still meaningful with the pipeline log off: it decides whether HY's
+    # DEBUG lines in hy_memory.log carry full model prompts and responses.
     os.environ["MEMORY_LOG_LLM_CONTENT"] = "true" if level == "debug" else "false"
     os.environ["MEMORY_MEMORY_OPERATIONS_ENABLED"] = "false"
     os.environ["ANONYMIZED_TELEMETRY"] = "False"

memhub/apps/core/src/one_memory/hy_sdk_migration.py

@@ -426,6 +426,7 @@ def _repair_coding(
         path = root / name
         if not path.is_file():
             continue
+        _ensure_coding_schema(path)
         with closing(sqlite3.connect(path)) as database:
             tables = {
                 str(row[0])
@@ -437,10 +438,14 @@ def _repair_coding(
             ):
                 if table not in tables:
                     continue
+                # _ensure_coding_schema added it; an SDK without the column
+                # (before 2.0.7) leaves the time to the vector repair alone.
+                has_memory_at = "memory_at" in _columns(database, table)
                 rows = database.execute(
-                    f"SELECT {identifier},source,created_at FROM {table}"
+                    f"SELECT {identifier},source,created_at,"
+                    f"{'memory_at' if has_memory_at else 'NULL'} FROM {table}"
                 ).fetchall()
-                for memory_id, source, created_at in rows:
+                for memory_id, source, created_at, memory_at in rows:
                     desired = mapping.get(str(memory_id))
                     if desired is None:
                         continue
@@ -449,6 +454,13 @@ def _repair_coding(
                     if desired.source and source in {None, "", "default"}:
                         updates.append("source=?")
                         values.append(desired.source)
+                    if (
+                        has_memory_at
+                        and desired.memory_at is not None
+                        and memory_at != desired.memory_at
+                    ):
+                        updates.append("memory_at=?")
+                        values.append(desired.memory_at)
                     if desired.created_at is not None:
                         desired_created_at = (
                             datetime.fromtimestamp(desired.created_at, tz=UTC)
@@ -470,6 +482,32 @@ def _repair_coding(
     return repaired
 
 
+def _ensure_coding_schema(path: Path) -> None:
+    """Bring a Coding/Working store to the installed SDK's schema.
+
+    A migrated home's stores were written by an older SDK and may lack columns
+    the repair writes (``memory_at`` since hy-memory 2.0.7). The repair runs
+    before any SDK client opens them, so the SDK's own schema step is run here
+    rather than repeating its DDL. ``CodingMemoryStore`` has no public
+    schema-only entry point: ``initialize()`` also opens its Chroma
+    collections, which must stay closed during the offline repair. So only its
+    SQLite step runs, on a store object that carries nothing but that path.
+    """
+    from hy_memory.coding.store import (  # type: ignore[import-untyped]
+        CodingMemoryStore,
+    )
+
+    # Relies on the pinned SDK's internals (hy-memory==2.0.9 is pinned exactly).
+    store = CodingMemoryStore.__new__(CodingMemoryStore)
+    store._db_path = str(path)
+    store._conn = None
+    try:
+        store._init_sqlite_sync()
+    finally:
+        if store._conn is not None:
+            store._conn.close()
+
+
 def _batches[T](values: list[T] | tuple[T, ...]) -> Iterator[tuple[T, ...]]:
     for start in range(0, len(values), _BATCH):
         yield tuple(values[start : start + _BATCH])

memhub/apps/core/src/one_memory/hy_sdk_runtime.py

@@ -11,7 +11,7 @@ import asyncio
 import hashlib
 import os
 import threading
-from collections.abc import Callable, Coroutine
+from collections.abc import Callable, Coroutine, Iterable, Mapping
 from dataclasses import dataclass
 from pathlib import Path
 from typing import Any, ClassVar
@@ -31,11 +31,18 @@ from one_memory.embedding_assets import (
 )
 from one_memory.hy_graph_admin import (
     GraphWidthMismatchError,
-    graph_store_enabled,
+    effective_memory_mode,
     inspect_graph_width_mismatch,
 )
 from one_memory.hy_sdk_types import HySdkRuntimeError
 from one_memory.local_embedding import LocalEmbedder
+from one_memory.log_file import confine_hy_sdk_logging
+from one_memory.memory_features import (
+    FEATURES,
+    FEATURES_BY_NAME,
+    FeatureOverrides,
+    feature_overrides,
+)
 from one_memory.model_profiles import (
     ModelProfileSnapshot,
     ProviderKind,
@@ -66,6 +73,16 @@ _STORE_GATE = threading.Condition()
 _OFFLINE_STORES: set[Path] = set()
 _OPENING: dict[Path, int] = {}
 
+# One client constructs at a time per home, its first-time store setup
+# included. The constructor opens Chroma's `PersistentClient`, and two first
+# opens on one directory race over that same unlocked System cache. On a
+# clean install the write drain's client and a UI read client started in the
+# same second; neither reached `VectorStore ready`, and the first write job
+# was lost as unknown. Taken inside `acquire_client`, after the offline gate
+# and never while `_STORE_GATE` is held, so it cannot deadlock with
+# `hold_stores_offline`, which refuses while `_OPENING` counts a client.
+_CONSTRUCTING: dict[Path, threading.Lock] = {}
+
 
 class LocalOnnxEmbeddingService:
     """Public ``EmbeddingService`` implementation backed by MemHub's ONNX model."""
@@ -159,11 +176,17 @@ class SdkClientHandle:
         api_key: str | None,
         data_dir: Path,
         http_transport_factory: Callable[[], Any] | None = None,
+        mode: str | None = None,
+        features: FeatureOverrides = (),
     ) -> None:
         self._profile = profile
         self._api_key = api_key
         self._data_dir = data_dir
         self._http_transport_factory = http_transport_factory
+        #: The SDK mode this handle's client is built with, fixed for its life.
+        self.mode = mode or effective_memory_mode()
+        #: The SDK features set on its config (`memory_features`), fixed likewise.
+        self.features = features
         self._lease: ClientLease | None = None
         self._lock = asyncio.Lock()
 
@@ -178,6 +201,8 @@ class SdkClientHandle:
                     api_key=self._api_key,
                     data_dir=self._data_dir,
                     http_transport_factory=self._http_transport_factory,
+                    mode=self.mode,
+                    features=self.features,
                 )
             return self._lease.client
 
@@ -206,6 +231,33 @@ def builtin_retrieval_snapshot() -> ModelProfileSnapshot:
     )
 
 
+def retrieval_view(profile: ModelProfileSnapshot) -> ModelProfileSnapshot:
+    """The read side of `profile`: its embedding space, and no generation.
+
+    Reads must open the collection writes fill (#242), and the collection is
+    named by the embedding only. Generation stays the built-in placeholder, so
+    a read never needs the generation model. A remote embedding still needs
+    the profile's API key, which `build_sdk_config` uses as the embedding key.
+    """
+    provider = profile.embedding_provider_kind or profile.provider_kind
+    endpoint: str | None = None
+    if provider is not ProviderKind.LOCAL_ONNX:
+        endpoint = profile.embedding_base_url
+        if not endpoint:
+            raise HySdkRuntimeError("hy_sdk_model_endpoint_missing")
+    return builtin_retrieval_snapshot().model_copy(
+        update={
+            "profile_id": profile.profile_id,
+            "revision": profile.revision,
+            "embedding_provider_kind": provider,
+            "embedding_model": profile.embedding_model,
+            "embedding_base_url": endpoint,
+            "embedding_dimensions": profile.embedding_dimensions,
+            "has_api_key": profile.has_api_key,
+        }
+    )
+
+
 def sdk_relevance_floor() -> float:
     raw = os.environ.get("ONE_MEMORY_HY_RELEVANCE_FLOOR", "").strip()
     if not raw:
@@ -217,13 +269,48 @@ def sdk_relevance_floor() -> float:
     return min(1.0, max(0.0, value))
 
 
+def sdk_feature_defaults() -> dict[str, bool]:
+    """Each `memory_features` feature's SDK default, as a fresh config has it.
+
+    That is the SDK's code default unless the process environment sets the
+    SDK's own variable (`FeatureSpec.env`), exactly what an unset feature
+    leaves on a client.
+    """
+    config = MemoryConfig()
+    return {
+        spec.name: bool(getattr(getattr(config, spec.section), spec.field)) for spec in FEATURES
+    }
+
+
+def apply_sdk_features(
+    config: MemoryConfig, features: Mapping[str, bool] | Iterable[tuple[str, bool]]
+) -> MemoryConfig:
+    """Set each feature a person chose on its `MemoryConfig` field.
+
+    Every exposed feature is a config field the SDK reads from the config at
+    call time, never from `os.environ`, so this is per client and
+    deterministic. Nothing here touches `vector_store`, so the collection
+    identity never depends on a feature.
+    """
+    values = dict(features.items() if isinstance(features, Mapping) else features)
+    for name, value in feature_overrides(values):
+        spec = FEATURES_BY_NAME[name]
+        setattr(getattr(config, spec.section), spec.field, value)
+    return config
+
+
 def build_sdk_config(
     profile: ModelProfileSnapshot,
     *,
     api_key: str | None,
     data_dir: Path,
+    features: Mapping[str, bool] | Iterable[tuple[str, bool]] = (),
 ) -> MemoryConfig:
-    """Map a model snapshot to SDK models, endpoints, backends, and local paths."""
+    """Map a model snapshot to SDK models, endpoints, backends, and local paths.
+
+    `features` (`memory_features`) are set on top; with none the config is
+    exactly what it was before the setting existed.
+    """
     generation_endpoint, generation_key = _generation_access(profile, api_key)
     embedding_provider = profile.embedding_provider_kind or profile.provider_kind
     embedding_service_is_local = embedding_provider is ProviderKind.LOCAL_ONNX
@@ -238,10 +325,8 @@ def build_sdk_config(
         embedding_key = None
         dimensions = profile.embedding_dimensions or BUILTIN_EMBEDDING_DIMENSIONS
     else:
-        embedding_endpoint = _openai_endpoint(
-            embedding_provider,
-            profile.embedding_base_url or profile.base_url,
-        )
+        # Never the generation endpoint: the collection name hashes this one.
+        embedding_endpoint = _openai_endpoint(embedding_provider, profile.embedding_base_url)
         embedding_key = "ollama" if embedding_provider is ProviderKind.OLLAMA else api_key
         if embedding_provider is ProviderKind.OPENAI_COMPATIBLE and not embedding_key:
             raise HySdkRuntimeError("hy_sdk_embedding_api_key_missing")
@@ -271,7 +356,7 @@ def build_sdk_config(
     if thinking_off:
         llm["extra_body"] = thinking_off
 
-    return MemoryConfig.from_dict(
+    config = MemoryConfig.from_dict(
         {
             "llm": llm,
             "embedder": {
@@ -308,6 +393,7 @@ def build_sdk_config(
             "coding": {"db_path": str(root / "coding.sqlite3")},
         }
     )
+    return apply_sdk_features(config, features)
 
 
 def acquire_client(
@@ -316,9 +402,15 @@ def acquire_client(
     api_key: str | None,
     data_dir: Path,
     http_transport_factory: Callable[[], Any] | None = None,
+    mode: str | None = None,
+    features: FeatureOverrides = (),
 ) -> ClientLease:
     """Construct a public SDK client and return its explicit runtime lease.
 
+    `mode` is the SDK pipeline mode (lite / pro / ultra); None keeps the
+    pre-setting default (ultra when `ONE_MEMORY_GRAPH_STORE` is set, else pro).
+    `features` are the `memory_features` set on its config; none keeps every
+    SDK default.
     Waits while `hold_stores_offline` holds this home's stores.
     """
     key = data_dir.resolve()
@@ -332,6 +424,8 @@ def acquire_client(
             api_key=api_key,
             data_dir=data_dir,
             http_transport_factory=http_transport_factory,
+            mode=mode or effective_memory_mode(),
+            features=features,
         )
     finally:
         with _STORE_GATE:
@@ -370,8 +464,10 @@ def _acquire_client(
     api_key: str | None,
     data_dir: Path,
     http_transport_factory: Callable[[], Any] | None,
+    mode: str = "pro",
+    features: FeatureOverrides = (),
 ) -> ClientLease:
-    config = build_sdk_config(profile, api_key=api_key, data_dir=data_dir)
+    config = build_sdk_config(profile, api_key=api_key, data_dir=data_dir, features=features)
     config.http_transport_factory = http_transport_factory
     mismatch = inspect_graph_width_mismatch(config)
     if mismatch is not None:
@@ -381,16 +477,23 @@ def _acquire_client(
     embedding_service: LocalOnnxEmbeddingService | None = None
     if (profile.embedding_provider_kind or profile.provider_kind) is ProviderKind.LOCAL_ONNX:
         embedding_service = LocalOnnxEmbeddingService()
+    with _STORE_GATE:
+        constructing = _CONSTRUCTING.setdefault(key, threading.Lock())
     try:
-        client = HyMemoryClient(
-            config=config,
-            mode="ultra" if graph_store_enabled() else "pro",
-            runtime=runtime,
-            embedding_service=embedding_service,
-        )
+        with constructing:
+            client = HyMemoryClient(
+                config=config,
+                mode=mode,
+                runtime=runtime,
+                embedding_service=embedding_service,
+            )
     except BaseException:
         _release_runtime(key)
         raise
+    finally:
+        # The constructor's first step configures HY's text log (once per
+        # process), so by here it exists whether or not construction succeeded.
+        confine_hy_sdk_logging()
     return ClientLease(client, key)
 
 
@@ -449,7 +552,7 @@ def vector_collection_bases(profile: ModelProfileSnapshot) -> tuple[str, tuple[s
         endpoint, dimensions = None, profile.embedding_dimensions or BUILTIN_EMBEDDING_DIMENSIONS
         legacy_endpoint = "local-onnx"
     else:
-        endpoint = _openai_endpoint(provider, profile.embedding_base_url or profile.base_url)
+        endpoint = _openai_endpoint(provider, profile.embedding_base_url)
         dimensions = profile.embedding_dimensions
         legacy_endpoint = endpoint
     legacy_model = profile.embedding_model or "bge-small-zh-v1.5"

memhub/apps/core/src/one_memory/hy_sdk_types.py

@@ -2,6 +2,7 @@
 
 from __future__ import annotations
 
+import re
 from collections.abc import Mapping
 from datetime import UTC, datetime
 from typing import Literal, NamedTuple, cast
@@ -151,6 +152,7 @@ class HySdkMemoryRecord(BaseModel):
         if not content.strip():
             raise HySdkReadError("hy_memory_record_shape_unknown")
         status = str(raw.get("status") or "active")
+        memory_at = raw.get("memory_at")
         created_at = _epoch(raw.get("created_at"))
         updated_at = _epoch(raw.get("updated_at"))
         try:
@@ -167,6 +169,9 @@ class HySdkMemoryRecord(BaseModel):
                     str(raw["session_id"]) if isinstance(raw.get("session_id"), str) else None
                 ),
                 source=str(raw["source"]) if isinstance(raw.get("source"), str) else None,
+                # Conversation time; hy-memory 2.0.7 stores it for experiences
+                # and facts. Older stores leave it unknown.
+                memory_at=memory_at if isinstance(memory_at, int) else None,
                 gmt_created=created_at,
                 scene=cast(MemoryScene, variant),
                 kind=cast(MemoryKindName, kind),
@@ -387,6 +392,8 @@ class HySdkRawRecord(BaseModel):
     memory_at: int | None = None
     gmt_created: int | None = None
     custom: dict[str, object] = Field(default_factory=dict)
+    #: active, or shadow once pro extracted it; None when the SDK gave none.
+    status: str | None = None
 
     @classmethod
     def from_sdk(cls, raw: object) -> HySdkRawRecord:
@@ -413,11 +420,34 @@ class HySdkRawRecord(BaseModel):
                     raw.get("gmt_created") if isinstance(raw.get("gmt_created"), int) else None
                 ),
                 custom=dict(custom),
+                status=str(raw["status"]) if isinstance(raw.get("status"), str) else None,
             )
         except (KeyError, TypeError, ValueError) as error:
             raise HySdkReadError("hy_memory_raw_shape_unknown") from error
 
 
+_RAW_TURN_LINE = re.compile(r"^\[(user|assistant|tool|system)\]:\s?(.*)$")
+
+
+def raw_turn_messages(content: str) -> tuple[tuple[str, str], ...]:
+    """The `(role, text)` messages of one SDK raw turn.
+
+    The SDK stores a turn as `[user]: …\\n[assistant]: …`; a line without a
+    role prefix continues the message above it. Content that names no role
+    at all is one message of unknown role (`""`).
+    """
+    messages: list[list[str]] = []
+    for line in content.splitlines():
+        match = _RAW_TURN_LINE.match(line)
+        if match is not None:
+            messages.append([match.group(1), match.group(2)])
+        elif messages:
+            messages[-1][1] += "\n" + line
+        else:
+            messages.append(["", line])
+    return tuple((role, text) for role, text in messages)
+
+
 class HySdkRawHit(HySdkRawRecord):
     score: float = Field(ge=0, le=1)
 
@@ -512,6 +542,96 @@ class HySdkUserProfile(BaseModel):
         )
 
 
+class HySdkProfileVersion(BaseModel):
+    """One version on an L0 basic-info chain."""
+
+    model_config = ConfigDict(frozen=True, extra="forbid")
+
+    memory_id: str
+    content: str
+    #: Conversation time as the SDK renders it on a chain (epoch or text).
+    memory_at: int | str | None = None
+    gmt_created: int | None = None
+    #: The chain head: the version the profile shows now.
+    current: bool = False
+
+
+class HySdkProfileChain(BaseModel):
+    """One L0 basic-info head and its versions, latest first."""
+
+    model_config = ConfigDict(frozen=True, extra="forbid")
+
+    memory_id: str
+    content: str
+    status: str
+    memory_at: int | None = None
+    gmt_created: int | None = None
+    #: The structured fields this head holds (`basic_info_kv`).
+    basic_info: dict[str, object] = Field(default_factory=dict)
+    versions: tuple[HySdkProfileVersion, ...]
+
+
+class HySdkProfileVersions(BaseModel):
+    """`get_user_basic_info`: the long-term profile's L0 version chains."""
+
+    model_config = ConfigDict(frozen=True, extra="forbid")
+
+    request_id: str
+    chains: tuple[HySdkProfileChain, ...]
+    elapsed_ms: float = Field(ge=0)
+
+    @classmethod
+    def from_sdk(cls, raw: object) -> HySdkProfileVersions:
+        if not isinstance(raw, Mapping):
+            raise HySdkReadError("hy_memory_profile_versions_shape_unknown")
+        request_id = raw.get("request_id")
+        rows = raw.get("memories")
+        elapsed_ms = raw.get("elapsed_ms")
+        if (
+            not isinstance(request_id, str)
+            or not isinstance(rows, list)
+            or not isinstance(elapsed_ms, int | float)
+            or isinstance(elapsed_ms, bool)
+        ):
+            raise HySdkReadError("hy_memory_profile_versions_shape_unknown")
+        chains: list[HySdkProfileChain] = []
+        for row in rows:
+            if not isinstance(row, Mapping) or not isinstance(row.get("memory_id"), str):
+                raise HySdkReadError("hy_memory_profile_versions_shape_unknown")
+            head_id = str(row["memory_id"])
+            chain = row.get("evolution_chain")
+            entries = chain if isinstance(chain, list) and chain else [{**row, "node_id": head_id}]
+            versions = tuple(
+                HySdkProfileVersion(
+                    memory_id=str(entry.get("node_id") or entry.get("memory_id")),
+                    content=str(entry.get("content") or ""),
+                    memory_at=(
+                        entry.get("memory_at")
+                        if isinstance(entry.get("memory_at"), int | str)
+                        and not isinstance(entry.get("memory_at"), bool)
+                        else None
+                    ),
+                    gmt_created=_epoch(entry.get("gmt_created")),
+                    current=str(entry.get("node_id") or entry.get("memory_id")) == head_id,
+                )
+                for entry in entries
+                if isinstance(entry, Mapping) and (entry.get("node_id") or entry.get("memory_id"))
+            )
+            basic_info = row.get("basic_info_kv")
+            chains.append(
+                HySdkProfileChain(
+                    memory_id=head_id,
+                    content=str(row.get("content") or ""),
+                    status=str(row.get("status") or "active"),
+                    memory_at=_epoch(row.get("memory_at")),
+                    gmt_created=_epoch(row.get("gmt_created")),
+                    basic_info=dict(basic_info) if isinstance(basic_info, Mapping) else {},
+                    versions=versions,
+                )
+            )
+        return cls(request_id=request_id, chains=tuple(chains), elapsed_ms=float(elapsed_ms))
+
+
 def parse_sdk_search(raw: object, *, error_code: str) -> tuple[SemanticMemoryHit, ...]:
     rows = _reply_rows(raw, error_code)
     found: list[SemanticMemoryHit] = []
@@ -538,6 +658,120 @@ def parse_sdk_search(raw: object, *, error_code: str) -> tuple[SemanticMemoryHit
     return tuple(found)
 
 
+def parse_sdk_lane_search(raw: object, *, error_code: str) -> tuple[SemanticMemoryHit, ...]:
+    """Every scene's search reply as hits that name their lane and namespace.
+
+    `parse_sdk_search` reads the chat reply only and forgets which lane a hit
+    came from. A caller may now pick the scene and a floor per lane, so this
+    reads all three reply shapes:
+
+    - chat: `memories` as `{schema, proactive, normal}` (or one list, normal);
+    - working/coding: `facts` and `experiences` of that variant;
+    - all: one fused `memories` list whose rows carry `scene`, `kind` and
+      their own lane's score (`lane_score`). The fused `score` is a rank
+      fusion (1/(60+rank)), not a relevance, so the lane score is reported.
+
+    Highest score first, first occurrence of an id kept.
+    """
+    if not isinstance(raw, Mapping):
+        raise HySdkReadError(error_code)
+    scene = raw.get("scene")
+    found: list[SemanticMemoryHit] = []
+    if scene == "all":
+        rows = raw.get("memories")
+        if not isinstance(rows, list) or any(not isinstance(row, Mapping) for row in rows):
+            raise HySdkReadError(error_code)
+        for row in rows:
+            hit = _lane_hit(row, scene=str(row.get("scene") or "chat"), score_key="lane_score")
+            if hit is not None:
+                found.append(hit)
+    elif scene in ("productivity", "working", "coding"):
+        variant = raw.get("variant") if raw.get("variant") in {"working", "coding"} else None
+        lane = variant or (scene if scene in {"working", "coding"} else "working")
+        for key, kind in (("facts", "fact"), ("experiences", "experience")):
+            rows = raw.get(key) or []
+            if not isinstance(rows, list) or any(not isinstance(row, Mapping) for row in rows):
+                raise HySdkReadError(error_code)
+            for row in rows:
+                hit = _lane_hit({**row, "kind": kind}, scene=lane, score_key="score")
+                if hit is not None:
+                    found.append(hit)
+    else:
+        memories = raw.get("memories")
+        if isinstance(memories, Mapping):
+            if set(memories) != {"schema", "proactive", "normal"}:
+                raise HySdkReadError(error_code)
+            channels = tuple(memories.items())
+        elif isinstance(memories, list):
+            channels = (("normal", memories),)
+        else:
+            raise HySdkReadError(error_code)
+        for lane, rows in channels:
+            if not isinstance(rows, list) or any(not isinstance(row, Mapping) for row in rows):
+                raise HySdkReadError(error_code)
+            for row in rows:
+                hit = _lane_hit(row, scene="chat", score_key="score", lane=str(lane))
+                if hit is not None:
+                    found.append(hit)
+    ordered: list[SemanticMemoryHit] = []
+    seen: set[tuple[str, str, str | None]] = set()
+    for hit in sorted(found, key=lambda value: value.score, reverse=True):
+        identity = (hit.memory_id, hit.kind, hit.variant)
+        if identity in seen:
+            continue
+        seen.add(identity)
+        ordered.append(hit)
+    return tuple(ordered)
+
+
+def _lane_hit(
+    row: Mapping[str, object],
+    *,
+    scene: str,
+    score_key: str,
+    lane: str | None = None,
+) -> SemanticMemoryHit | None:
+    kind = str(row.get("kind")) if row.get("kind") in {"experience", "fact"} else "chat"
+    if kind == "chat":
+        memory_id = row.get("memory_id") or row.get("id")
+        variant = None
+        hit_lane = lane or "normal"
+    else:
+        id_key = "experience_id" if kind == "experience" else "fact_id"
+        memory_id = row.get("memory_id") or row.get(id_key)
+        variant = scene if scene in {"working", "coding"} else None
+        if variant is None:
+            return None
+        hit_lane = variant
+    if not isinstance(memory_id, str) or not memory_id:
+        return None
+    value = row.get(score_key)
+    if not isinstance(value, int | float) or isinstance(value, bool):
+        value = row.get("score")
+    score = (
+        max(0.0, min(1.0, float(value)))
+        if isinstance(value, int | float) and not isinstance(value, bool)
+        else 0.0
+    )
+    if kind != "chat":
+        return SemanticMemoryHit(
+            memory_id=memory_id, score=score, lane=hit_lane, kind=kind, variant=variant
+        )
+    content = row.get("content")
+    member_ids, formatting_complete = _formatted_member_ids(row)
+    return SemanticMemoryHit(
+        memory_id=memory_id,
+        score=score,
+        formatted_content=(
+            content.strip() if isinstance(content, str) and content.strip() else None
+        ),
+        member_ids=member_ids,
+        sdk_formatted=True,
+        formatting_complete=formatting_complete,
+        lane=hit_lane,
+    )
+
+
 class LibraryHit(NamedTuple):
     """One keyword match in a library search window, with its namespace."""
 

memhub/apps/core/src/one_memory/imports/adapters.py

@@ -31,6 +31,16 @@ class ScanIssueCode(StrEnum):
     #: import it. Named for the reason, so a reader can tell a setting from
     #: a record the parser could not use.
     SUBAGENT_TRACE_DISABLED = "subagent_trace_disabled"
+    #: A host's own notice wearing the user's role — Hermes posts a finished
+    #: background process, and its gateway marks the post as not written by a
+    #: human. Nothing in it is the person's, so it never reaches extraction.
+    INJECTED_MACHINE_NOTICE = "injected_machine_notice"
+    #: The host could not transcribe a voice message and left a placeholder
+    #: naming the audio file. The person did speak, but the placeholder holds
+    #: only a path, which an extractor would otherwise read as a fact.
+    VOICE_PLACEHOLDER = "voice_placeholder"
+    #: A health check such as "Reply with exactly: PONG": no one's words.
+    PROBE = "probe"
     SOURCE_BUSY = "source_busy"
     UNSUPPORTED_SCHEMA = "unsupported_schema"
     BUDGET_EXCEEDED = "budget_exceeded"

memhub/apps/core/src/one_memory/imports/discovery.py

@@ -27,6 +27,7 @@ def discovered_off_loop(
     *,
     suffixes: frozenset[str],
     max_targets: int,
+    excluded_top_directories: frozenset[str] = frozenset(),
 ) -> AsyncIterator[SourceFileDiscoveryEvent]:
     """The same walk, for callers that have an event loop to keep free.
 
@@ -41,6 +42,7 @@ def discovered_off_loop(
             grant,
             suffixes=suffixes,
             max_targets=max_targets,
+            excluded_top_directories=excluded_top_directories,
         ),
         scope_key=grant.source_id,
     )
@@ -51,12 +53,15 @@ def discover_source_files(
     *,
     suffixes: frozenset[str],
     max_targets: int,
+    excluded_top_directories: frozenset[str] = frozenset(),
 ) -> Iterator[SourceFileDiscoveryEvent]:
     """Metadata only: one lstat per candidate, no open and no per-file grant lookup.
 
     The grant and root identity are checked once before the walk. Each listed
     file keeps the grant's check, which runs again when that file is opened;
     the open itself refuses symlinks and any inode other than the one listed.
+    `excluded_top_directories` are directories directly under the root that
+    are never entered, such as another user's Hermes profiles.
     """
     grant.check_root()
     inspected = 0
@@ -64,8 +69,12 @@ def discover_source_files(
         grant.root,
         followlinks=False,
     ):
+        at_root = Path(directory) == grant.root
         child_directories[:] = sorted(
-            child for child in child_directories if not (Path(directory) / child).is_symlink()
+            child
+            for child in child_directories
+            if not (Path(directory) / child).is_symlink()
+            and not (at_root and child in excluded_top_directories)
         )
         for filename in sorted(filenames):
             if Path(filename).suffix.lower() not in suffixes:

memhub/apps/core/src/one_memory/imports/grants.py

@@ -252,6 +252,18 @@ class SourceGrantStore:
             ).fetchone()
         return self._to_receipt(row) if row else None
 
+    def grants(self) -> tuple[tuple[SourceGrantReceipt, str], ...]:
+        """Every grant of every user, with its granted root path.
+
+        One source may hold several grants, one per `(user_id, source_id)`:
+        a Hermes root for `local` and a Hermes profile for another user.
+        """
+        with self._connect() as connection:
+            rows = connection.execute(
+                "SELECT * FROM source_grants ORDER BY source_id, user_id"
+            ).fetchall()
+        return tuple((self._to_receipt(row), str(row["requested_path"])) for row in rows)
+
     def require_receipt(self, *, user_id: str, source_id: str) -> SourceGrantReceipt:
         receipt = self.receipt(user_id=user_id, source_id=source_id)
         if receipt is None:

memhub/apps/core/src/one_memory/imports/live_sync.py

@@ -28,6 +28,7 @@ from one_memory.imports.adapters import (
 )
 from one_memory.imports.jobs import ImportJob, ImportJobStore, ImportStatus
 from one_memory.imports.supervisor import ImportSupervisor
+from one_memory.user_scope import DEFAULT_USER_ID, lane_key, split_lane
 
 #: How long an automatic retry will wait for a wedged worker to notice cancel.
 ABANDON_TIMEOUT_SECONDS = 5.0
@@ -64,7 +65,12 @@ class LiveSyncHealth(BaseModel):
 
 
 class PollingLiveSyncSupervisor:
-    """Near-live sync that reuses each adapter's persisted incremental cursor."""
+    """Near-live sync that reuses each adapter's persisted incremental cursor.
+
+    A "source" here is a lane key (`user_scope.lane_key`): the bare source id
+    for `local`, `source@user` for another user's grant. Cursors, jobs and
+    locks are therefore per `(user_id, source_id)`.
+    """
 
     def __init__(
         self,
@@ -92,8 +98,22 @@ class PollingLiveSyncSupervisor:
         }
         self._locks = {source_id: asyncio.Lock() for source_id in self.source_ids}
         self._task: asyncio.Task[None] | None = None
+        self._group: asyncio.TaskGroup | None = None
         self._observers: set[asyncio.Task[None]] = set()
 
+    def ensure_source(self, source_id: str) -> None:
+        """Follow a lane granted after start, such as a new Hermes profile."""
+        if source_id in self._locks:
+            return
+        self.source_ids = (*self.source_ids, source_id)
+        self._health.setdefault(
+            source_id,
+            LiveSyncHealth(source_id=source_id, mode=self.mode, state=LiveSyncState.IDLE),
+        )
+        self._locks[source_id] = asyncio.Lock()
+        if self._group is not None:
+            self._group.create_task(self._poll_source(source_id))
+
     def _set_health(self, source_id: str, value: LiveSyncHealth) -> None:
         previous = self._health.get(source_id)
         self._health[source_id] = value
@@ -152,7 +172,12 @@ class PollingLiveSyncSupervisor:
         self._watched_sources = frozenset(source_ids)
 
     def all_health(self) -> tuple[LiveSyncHealth, ...]:
-        return tuple(self.health(source_id) for source_id in self.source_ids)
+        """`local`'s lanes only: what the App has always been shown."""
+        return tuple(
+            self.health(source_id)
+            for source_id in self.source_ids
+            if split_lane(source_id)[0] == DEFAULT_USER_ID
+        )
 
     async def initialize_source(
         self, source_id: str, *, baseline_missing_files: bool = True
@@ -164,11 +189,12 @@ class PollingLiveSyncSupervisor:
         baselines = (
             await adapter.live_baseline() if isinstance(adapter, _LiveBaselineAdapter) else {}
         )
+        user_id, bare_source_id = split_lane(source_id)
         await in_the_background(
             partial(
                 self.job_store.ensure_live_cursors,
-                user_id="local",
-                source_id=source_id,
+                user_id=user_id,
+                source_id=bare_source_id,
                 baselines=baselines,
                 baseline_missing_files=baseline_missing_files,
             )
@@ -228,7 +254,7 @@ class PollingLiveSyncSupervisor:
                 (
                     job
                     for job in reversed(source_jobs)
-                    if job.source_id == source_id
+                    if lane_key(job.user_id, job.source_id) == source_id
                     and job.operation == "sync"
                     and job.status
                     in {
@@ -277,13 +303,13 @@ class PollingLiveSyncSupervisor:
             competing = [
                 job
                 for job in source_jobs
-                if job.source_id == source_id
+                if lane_key(job.user_id, job.source_id) == source_id
                 and job.operation != "sync"
                 and job.status
                 in {ImportStatus.PENDING, ImportStatus.RUNNING, ImportStatus.INTERRUPTED}
             ]
             bounded_history_incomplete = any(
-                job.source_id == source_id
+                lane_key(job.user_id, job.source_id) == source_id
                 and job.operation == "history"
                 and job.history_after is not None
                 and job.status is not ImportStatus.COMPLETED
@@ -379,11 +405,12 @@ class PollingLiveSyncSupervisor:
                 await close()
 
     async def _live_cursors(self, source_id: str, adapter: object) -> dict[str, SourceCursor]:
+        user_id, bare_source_id = split_lane(source_id)
         cursors = await in_the_background(
             partial(
                 self.job_store.get_source_cursors,
-                user_id="local",
-                source_id=source_id,
+                user_id=user_id,
+                source_id=bare_source_id,
                 operation="sync",
             )
         )
@@ -395,8 +422,8 @@ class PollingLiveSyncSupervisor:
         return await in_the_background(
             partial(
                 self.job_store.ensure_live_cursors,
-                user_id="local",
-                source_id=source_id,
+                user_id=user_id,
+                source_id=bare_source_id,
                 baselines=baselines,
             )
         )
@@ -520,9 +547,13 @@ class PollingLiveSyncSupervisor:
     async def _run(self) -> None:
         # A source owns its cadence; one large catch-up cannot delay every
         # other source. The group keeps shutdown responsible for every poller.
-        async with asyncio.TaskGroup() as group:
-            for source_id in self.source_ids:
-                group.create_task(self._poll_source(source_id))
+        try:
+            async with asyncio.TaskGroup() as group:
+                self._group = group
+                for source_id in self.source_ids:
+                    group.create_task(self._poll_source(source_id))
+        finally:
+            self._group = None
 
     async def _poll_source(self, source_id: str) -> None:
         while True:

memhub/apps/core/src/one_memory/imports/runner.py

@@ -3,7 +3,7 @@ import logging
 import math
 import sqlite3
 from collections import Counter
-from collections.abc import Callable
+from collections.abc import AsyncIterator, Callable
 from contextlib import suppress
 from contextvars import ContextVar
 from dataclasses import dataclass, field
@@ -227,14 +227,17 @@ class ImportRunner:
         heartbeat = asyncio.create_task(self._keep_alive(job_id=job_id, runner_token=runner_token))
         pending = _PendingSkips()
         writes = _PendingWrites()
-        stream = adapter.scan(
-            ScanOptions(
-                cursors=cursors,
-                reuse_listing=(
-                    job.operation == "history"
-                    or (job.operation == "sync" and job.source_id == "workbuddy")
-                ),
-            )
+        stream = _owned_by(
+            adapter.scan(
+                ScanOptions(
+                    cursors=cursors,
+                    reuse_listing=(
+                        job.operation == "history"
+                        or (job.operation == "sync" and job.source_id == "workbuddy")
+                    ),
+                )
+            ),
+            job.user_id,
         )
         slice_cursors = dict(cursors)
         next_event: asyncio.Future[ScanEvent] | None = None
@@ -924,3 +927,30 @@ async def _settled_work[T](
         with suppress(asyncio.CancelledError, Exception):
             task.result()
         raise
+
+
+def _owned_by(stream: AsyncIterator[ScanEvent], user_id: str) -> AsyncIterator[ScanEvent]:
+    """Stamp every scanned record with the user whose root is being read.
+
+    Adapters know nothing of users and emit `local`. Which user a record
+    belongs to is decided by the grant the job reads (a Hermes profile's
+    `state.db` is that profile's user), so the job's user is the record's.
+    `local` passes the stream through untouched.
+    """
+    if user_id == "local":
+        return stream
+    return _restamped(stream, user_id)
+
+
+async def _restamped(stream: AsyncIterator[ScanEvent], user_id: str) -> AsyncIterator[ScanEvent]:
+    try:
+        async for event in stream:
+            if isinstance(event, ScannedEvidence) and event.evidence.user_id != user_id:
+                event = event.model_copy(
+                    update={"evidence": event.evidence.model_copy(update={"user_id": user_id})}
+                )
+            yield event
+    finally:
+        close = getattr(stream, "aclose", None)
+        if close is not None:
+            await close()

memhub/apps/core/src/one_memory/imports/sources/hermes.py

@@ -1,4 +1,5 @@
 import json
+import re
 import sqlite3
 from collections.abc import AsyncIterator, Iterator
 from functools import partial
@@ -87,6 +88,7 @@ class HermesSourceAdapter:
             self.grant,
             suffixes=frozenset({".jsonl", ".db"}),
             max_targets=10_000,
+            excluded_top_directories=frozenset({HERMES_PROFILES_DIRECTORY}),
         ):
             if isinstance(item, AuthorizedSourceFile) and _is_hermes_target(
                 item.relative_path.as_posix()
@@ -100,6 +102,7 @@ class HermesSourceAdapter:
             self.grant,
             suffixes=frozenset({".jsonl", ".db"}),
             max_targets=options.max_scan_targets or 100_000,
+            excluded_top_directories=frozenset({HERMES_PROFILES_DIRECTORY}),
         ):
             yield ScanProgressEvent(
                 phase=ScanPhase.DISCOVER,
@@ -214,7 +217,11 @@ def parse_hermes_rollout_record(
         message.get("content"),
         allowed_block_types=frozenset({"text", "input_text", "output_text"}),
     )
-    text = strip_injected_memory(text)
+    text = strip_injected_memory(text or "")
+    if role == "user":
+        text, injected = gateway_user_text(text)
+        if injected is not None:
+            return ParsedSkip(reason_code=injected, conversation_id=conversation_id)
     if not text:
         return ParsedSkip(
             reason_code=ScanIssueCode.NO_TEXT,
@@ -242,8 +249,17 @@ def parse_hermes_rollout_record(
     )
 
 
+#: Hermes keeps each extra profile under `<root>/profiles/<name>/`, with its
+#: own `state.db` and sessions. A profile is another person's history: it is
+#: read only through its own grant (its own MemHub user), never as part of the
+#: root above it. Hermes' own export skips this directory for the same reason.
+HERMES_PROFILES_DIRECTORY = "profiles"
+
+
 def _is_hermes_target(relative_path: str) -> bool:
     path = PurePosixPath(relative_path)
+    if path.parts and path.parts[0] == HERMES_PROFILES_DIRECTORY:
+        return False
     return path.name == "state.db" or (path.suffix == ".jsonl" and "sessions" in path.parts)
 
 
@@ -389,6 +405,15 @@ def _read_state_db(
             # `tool` rows after it hold only what came back.
             calls = _tool_calls_text(row["tool_calls"]) if role == "assistant" else None
             content = strip_injected_memory(str(row["content"] or ""))
+            if role == "user":
+                content, injected = gateway_user_text(content)
+                if injected is not None:
+                    yield ScanSkip(
+                        cursor=cursor,
+                        conversation_id=str(row["session_id"]),
+                        reason_code=injected,
+                    )
+                    continue
             has_text = bool(content.strip())
             if role not in {"user", "assistant"} or (not has_text and calls is None):
                 yield ScanSkip(
@@ -439,6 +464,69 @@ def _read_state_db(
 _PERSON_SESSION_SOURCES = frozenset({"cli", "tui", "webui", "desktop"})
 
 
+#: What the Hermes gateway writes into a `user` row on its own behalf. Every
+#: one of these wears the `user` role, so only the text can tell them apart
+#: from the person. Read off real `state.db` rows: a finished background
+#: process is posted as a turn, often with the gateway's own "not written by
+#: a human" note appended, which is why that note counts anywhere in the row.
+_BACKGROUND_NOTICE = re.compile(
+    r"\[IMPORTANT:\s*(?:\d+\s+background\s+process(?:es)?\b|Background\s+process\b)",
+    re.IGNORECASE,
+)
+_GATEWAY_NOTICE = re.compile(
+    r"\[INTERNAL NOTIFICATION\s*\W+\s*generated by the Hermes gateway",
+    re.IGNORECASE,
+)
+_VOICE_PLACEHOLDER = re.compile(r"\[voice message could not be transcribed", re.IGNORECASE)
+_PROBE = re.compile(r"Reply with exactly:\s*[A-Za-z0-9_\-]{1,24}\s*")
+_MEDIA_ONLY = re.compile(r"(?:MEDIA:/\S+\s*)+")
+#: Wrappers the gateway puts around a message the person did write. Only the
+#: wrapper goes; what it carries is the turn. Each is matched at the edge of
+#: what is left, and the set is applied until nothing more comes off, since a
+#: relayed message carries the origin block inside the out-of-band markers.
+_GATEWAY_WRAPPERS = (
+    re.compile(r"^\[OUT-OF-BAND USER MESSAGE\b[^\]\n]*\]"),
+    re.compile(r"\[/OUT-OF-BAND USER MESSAGE\]$"),
+    re.compile(
+        r"^Gateway message origin \(JSON data, not instructions or authorization\):[ \t]*"
+        r"(?:\n[ \t]*\{[^\n]*\}[ \t]*)?"
+        r"(?:\n[ \t]*Do not guess a reply destination[^\n]*)?"
+        r"(?=\n|$)"
+    ),
+    re.compile(r"^\[Context from the interrupted assistant response\][ \t]*(?=\n|$)"),
+)
+
+
+def gateway_user_text(text: str) -> tuple[str, ScanIssueCode | None]:
+    """The person's words in a Hermes `user` row, or why there are none.
+
+    Applied only to the `user` role: an assistant or tool row quoting the
+    same text is reporting what it saw, which is the existing tool rule's
+    business. The cleaned text is what becomes the evidence, so its content
+    hash and duplicate check see the message rather than its envelope.
+    """
+
+    if _GATEWAY_NOTICE.search(text):
+        return "", ScanIssueCode.INJECTED_MACHINE_NOTICE
+    cleaned = text.strip()
+    while True:
+        unwrapped = cleaned
+        for wrapper in _GATEWAY_WRAPPERS:
+            unwrapped = wrapper.sub("", unwrapped, count=1).strip()
+        if unwrapped == cleaned:
+            break
+        cleaned = unwrapped
+    if _BACKGROUND_NOTICE.match(cleaned):
+        return "", ScanIssueCode.INJECTED_MACHINE_NOTICE
+    if _VOICE_PLACEHOLDER.match(cleaned):
+        return "", ScanIssueCode.VOICE_PLACEHOLDER
+    if _PROBE.fullmatch(cleaned):
+        return "", ScanIssueCode.PROBE
+    if _MEDIA_ONLY.fullmatch(cleaned):
+        return "", ScanIssueCode.NO_TEXT
+    return cleaned, None
+
+
 def _tool_calls_text(value: object) -> str | None:
     """OpenAI-shaped `tool_calls`: `[{"function": {"name", "arguments"}}]`, stored as JSON."""
     if isinstance(value, str):

memhub/apps/core/src/one_memory/imports/supervisor.py

@@ -24,6 +24,7 @@ from one_memory.service_levels import (
     ServiceLevelScheduler,
     ServiceResource,
 )
+from one_memory.user_scope import lane_key, split_lane
 
 
 class UnknownSourceError(KeyError):
@@ -82,36 +83,45 @@ class ImportSupervisor:
         self._tasks: dict[str, asyncio.Task[None]] = {}
         self._drains: dict[str, asyncio.Task[None]] = {}
 
-    def register_adapter(self, adapter: SourceAdapter) -> None:
-        self.adapters[adapter.descriptor.source_id] = adapter
+    def register_adapter(self, adapter: SourceAdapter, *, user_id: str = "local") -> None:
+        """Adapters are keyed by lane: the source id for `local`, `source@user`
+        for another user's grant of the same source."""
+        self.adapters[lane_key(user_id, adapter.descriptor.source_id)] = adapter
 
     def unregister_adapter(self, source_id: str) -> None:
         self.adapters.pop(source_id, None)
 
+    @staticmethod
+    def lane_of(job: ImportJob) -> str:
+        return lane_key(job.user_id, job.source_id)
+
     def active_job_ids(self) -> frozenset[str]:
         return frozenset(job_id for job_id, task in self._tasks.items() if not task.done())
 
     def create_job(
         self, source_id: str, *, operation: Literal["history", "sync"] = "history"
     ) -> ImportJob:
+        """`source_id` is a lane key: another user's lane creates that user's job."""
         if self.source_enabled is not None and not self.source_enabled(source_id):
             raise SourceAccessError(SourceAccessErrorCode.GRANT_REQUIRED)
-        return self.job_store.create(user_id="local", source_id=source_id, operation=operation)
+        user_id, bare_source_id = split_lane(source_id)
+        return self.job_store.create(user_id=user_id, source_id=bare_source_id, operation=operation)
 
     def enqueue(self, job_id: str) -> ImportJob:
         if job_id in self._drains:
             raise InvalidJobTransitionError("import_cleanup_in_progress")
         job = self.job_store.require(job_id)
-        if job.source_id not in self.adapters:
-            raise UnknownSourceError(job.source_id)
-        if self.source_enabled is not None and not self.source_enabled(job.source_id):
+        lane = self.lane_of(job)
+        if lane not in self.adapters:
+            raise UnknownSourceError(lane)
+        if self.source_enabled is not None and not self.source_enabled(lane):
             return job
         existing = self._tasks.get(job_id)
         if existing is None or existing.done():
             self._tasks[job_id] = asyncio.create_task(
                 self._execute(
                     job_id,
-                    source_id=job.source_id,
+                    source_id=lane,
                     operation=job.operation,
                 )
             )
@@ -130,8 +140,10 @@ class ImportSupervisor:
         # pending history. Each resumed runner will yield after one bounded slice.
         for status in (ImportStatus.INTERRUPTED, ImportStatus.PENDING):
             for job in self.job_store.list_jobs(status=status):
-                if job.source_id in self.adapters:
-                    if self.source_enabled is not None and not self.source_enabled(job.source_id):
+                if self.lane_of(job) in self.adapters:
+                    if self.source_enabled is not None and not self.source_enabled(
+                        self.lane_of(job)
+                    ):
                         continue
                     if job.status is not ImportStatus.PENDING:
                         self.job_store.queue_for_resume(job.job_id)
@@ -227,7 +239,7 @@ class ImportSupervisor:
                     }:
                         return
                     if self.source_enabled is not None and not await in_the_background(
-                        self.source_enabled, job.source_id
+                        self.source_enabled, self.lane_of(job)
                     ):
                         return
                     runner = ImportRunner(
@@ -238,7 +250,7 @@ class ImportSupervisor:
                         sdk_queue=self.sdk_queue,
                         sdk_authority=self.sdk_authority,
                     )
-                    adapter = self.adapters[job.source_id]
+                    adapter = self.adapters[self.lane_of(job)]
                     if self.prepare_adapter is not None:
                         adapter = await self.prepare_adapter(job, adapter)
 
@@ -273,7 +285,7 @@ class ImportSupervisor:
                 and retry_attempt < self.max_history_retries
                 and (
                     self.source_enabled is None
-                    or await in_the_background(self.source_enabled, current.source_id)
+                    or await in_the_background(self.source_enabled, self.lane_of(current))
                 )
             ):
                 delay = self.retry_base_seconds * (2**retry_attempt)

memhub/apps/core/src/one_memory/library_search_window.py

@@ -7,9 +7,9 @@ between pages. The matches are kept here under an opaque ``window_id`` that
 the client passes back with the next offset.
 
 A window is bound to the filters it was searched with (query, source,
-statuses, scene). It expires after ``ttl_seconds`` without being read, and the
-store keeps at most ``max_windows`` windows, dropping the least recently read
-first.
+statuses, scene) and to the user whose library it searched. It expires after
+``ttl_seconds`` without being read, and the store keeps at most
+``max_windows`` windows, dropping the least recently read first.
 Only memory identities are kept, never content: each page is read fresh from
 the SDK, so an edit or a delete shows on the next page read.
 """
@@ -28,13 +28,24 @@ from one_memory.hy_sdk_types import LibraryHit
 DEFAULT_WINDOW_TTL_SECONDS = 300.0
 DEFAULT_MAX_WINDOWS = 32
 
-WindowKey = tuple[str, str | None, tuple[str, ...], str]
+#: (query, source, statuses, scene, user, other filters as sorted name/values pairs)
+WindowKey = tuple[
+    str, str | None, tuple[str, ...], str, str, tuple[tuple[str, tuple[str, ...]], ...]
+]
 
 
 def window_key(
-    query: str, source: str | None, statuses: list[str], scene: str = "all"
+    query: str,
+    source: str | None,
+    statuses: list[str],
+    scene: str = "all",
+    user_id: str = "local",
+    filters: dict[str, list[str]] | None = None,
 ) -> WindowKey:
-    return (query, source, tuple(sorted(set(statuses))), scene)
+    extra = tuple(
+        sorted((name, tuple(sorted(set(values)))) for name, values in (filters or {}).items())
+    )
+    return (query, source, tuple(sorted(set(statuses))), scene, user_id, extra)
 
 
 @dataclass(frozen=True)

memhub/apps/core/src/one_memory/log_file.py

@@ -20,8 +20,11 @@ def start_logging_to_file(data_dir: Path) -> Path:
     nowhere. That includes the loop watchdog: the one report that names what
     stalled the app was unreadable in exactly the situation it exists for.
 
-    The active diagnostic collection level controls Core, HY SDK, and provider
-    client records. Uvicorn keeps its request log on its own non-propagating
+    The active diagnostic collection level controls Core and provider client
+    records. The HY SDK keeps its own text log, hy_memory.log under the SDK's
+    log directory, and stops propagating here once its first client is built;
+    `confine_hy_sdk_logging` then forwards only its warnings and errors to
+    this file. Uvicorn keeps its request log on its own non-propagating
     loggers.
     """
     destination = log_path(data_dir)
@@ -49,3 +52,53 @@ def start_logging_to_file(data_dir: Path) -> Path:
     if root.level > logging.INFO or root.level == logging.NOTSET:
         root.setLevel(logging.INFO)
     return destination
+
+
+HY_SDK_LOGGER = "hy_memory"
+
+
+class _CoreLogForwarder(logging.Handler):
+    """Hands HY warnings to Core's own core.log handler, rotation and format."""
+
+    def __init__(self, target: logging.Handler) -> None:
+        super().__init__(level=logging.WARNING)
+        self.target = target
+
+    def emit(self, record: logging.LogRecord) -> None:
+        self.target.handle(record)
+
+
+def confine_hy_sdk_logging() -> None:
+    """Keep HY's text log in hy_memory.log, with only its problems in core.log.
+
+    HY's first client sets up hy_memory.log plus a stdout and a stderr stream
+    on the "hy_memory" logger, and stops it propagating to the root. Under
+    launchd those streams are service.log and service-error.log, so every SDK
+    line was written twice, and none reached core.log. Drop the streams and
+    forward WARNING and above to core.log, which the Debug panel reads; the
+    SDK keeps conversation text out of those levels.
+
+    Idempotent. Does nothing until HY has configured its logger, because a
+    propagating logger already reaches core.log through the root.
+    """
+    sdk = logging.getLogger(HY_SDK_LOGGER)
+    if sdk.propagate:
+        return
+    for handler in list(sdk.handlers):
+        if isinstance(handler, logging.StreamHandler) and not isinstance(
+            handler, logging.FileHandler
+        ):
+            sdk.removeHandler(handler)
+    if any(isinstance(handler, _CoreLogForwarder) for handler in sdk.handlers):
+        return
+    core = next(
+        (
+            handler
+            for handler in logging.getLogger().handlers
+            if isinstance(handler, RotatingFileHandler)
+            and Path(handler.baseFilename).name == "core.log"
+        ),
+        None,
+    )
+    if core is not None:
+        sdk.addHandler(_CoreLogForwarder(core))

memhub/apps/core/src/one_memory/memory_additions.py

@@ -11,10 +11,11 @@ import asyncio
 import sqlite3
 from collections.abc import Sequence
 from datetime import UTC, datetime
+from functools import partial
 from typing import Any, Literal, cast
 from uuid import UUID
 
-from fastapi import APIRouter, BackgroundTasks, HTTPException, Request
+from fastapi import APIRouter, BackgroundTasks, HTTPException, Query, Request
 from fastapi.params import Depends
 from pydantic import BaseModel, ConfigDict, Field, field_validator
 
@@ -34,6 +35,7 @@ from one_memory.memory_server.durable_queue import DurableWriteQueue
 from one_memory.memory_server.unified_entry import UnifiedWriteIntake
 from one_memory.memory_tasks import MemoryTaskStatus, MemoryTaskStore
 from one_memory.secrets import SecretStoreUnavailableError
+from one_memory.user_scope import DEFAULT_USER_ID, USER_ID_PATTERN, UserId
 
 SOURCE_ID = "one-memory-manual"
 MemoryAddState = Literal[
@@ -53,6 +55,8 @@ class MemoryAddInput(BaseModel):
     request_id: UUID
     text: str = Field(min_length=1, max_length=12_000)
     scope: Literal["all-apps"]
+    #: Whose memory this is. The App never sends it and writes `local`.
+    user_id: UserId = DEFAULT_USER_ID
 
     @field_validator("text")
     @classmethod
@@ -108,7 +112,11 @@ class MemoryAdditionStore:
 
     def save(self, addition: MemoryAddInput) -> EvidenceRecord:
         now = datetime.now(UTC)
+        if addition.user_id != DEFAULT_USER_ID and not self.sdk_authority:
+            # Legacy authority keeps one person's library only.
+            raise ValueError("memory_add_user_requires_sdk_authority")
         value = EvidenceInput(
+            user_id=addition.user_id,
             source_id=SOURCE_ID,
             conversation_id=f"note:{addition.request_id}",
             message_id=str(addition.request_id),
@@ -200,7 +208,7 @@ class MemoryAdditionStore:
                 )
         return self.require(value.evidence_id)
 
-    def require(self, evidence_id: str) -> EvidenceRecord:
+    def require(self, evidence_id: str, *, user_id: str = DEFAULT_USER_ID) -> EvidenceRecord:
         if self.sdk_authority:
             if self.write_queue is None:
                 raise KeyError(evidence_id)
@@ -219,11 +227,11 @@ class MemoryAdditionStore:
                 (item for item in payload.input_records if item.evidence_id == evidence_id),
                 None,
             )
-            if record is None:
+            if record is None or record.user_id != user_id:
                 raise KeyError(evidence_id)
             return record
         record = self.evidence.get(evidence_id)
-        if record is None or record.user_id != "local" or record.source_id != SOURCE_ID:
+        if record is None or record.user_id != user_id or record.source_id != SOURCE_ID:
             raise KeyError(evidence_id)
         metadata = record.origin_meta.get(EvidenceOrigin.MANUAL_IMPORT)
         if metadata is None or metadata.raw_meta.get("input_mode") != "direct_memory_note":
@@ -237,8 +245,8 @@ class MemoryAdditionStore:
             raise ValueError("memory_add_source_revoked")
         return record
 
-    def receipt(self, evidence_id: str) -> MemoryAddReceipt:
-        record = self.require(evidence_id)
+    def receipt(self, evidence_id: str, *, user_id: str = DEFAULT_USER_ID) -> MemoryAddReceipt:
+        record = self.require(evidence_id, user_id=user_id)
         if self.sdk_authority:
             assert self.write_queue is not None
             with main_database(self.write_queue.database_path).read() as database:
@@ -300,10 +308,15 @@ class MemoryAdditionStore:
             error=error,
         )
 
-    def receipt_for_request(self, request_id: UUID) -> MemoryAddReceipt:
+    def receipt_for_request(
+        self, request_id: UUID, *, user_id: str = DEFAULT_USER_ID
+    ) -> MemoryAddReceipt:
         identity = str(request_id)
         if self.sdk_authority:
+            # The Evidence id carries the user, so another user's request id
+            # names a different note and reads as not found.
             probe = EvidenceInput(
+                user_id=user_id,
                 source_id=SOURCE_ID,
                 conversation_id=f"note:{identity}",
                 message_id=identity,
@@ -316,16 +329,16 @@ class MemoryAdditionStore:
                     )
                 },
             )
-            return self.receipt(probe.evidence_id)
+            return self.receipt(probe.evidence_id, user_id=user_id)
         with main_database(self.evidence.database_path).read() as db:
             row = db.execute(
-                "SELECT evidence_id FROM evidence WHERE user_id='local' AND source_id=? "
+                "SELECT evidence_id FROM evidence WHERE user_id=? AND source_id=? "
                 "AND conversation_id=? AND message_id=?",
-                (SOURCE_ID, f"note:{identity}", identity),
+                (user_id, SOURCE_ID, f"note:{identity}", identity),
             ).fetchone()
         if row is None:
             raise KeyError(identity)
-        return self.receipt(str(row[0]))
+        return self.receipt(str(row[0]), user_id=user_id)
 
 
 def memory_add_router(dependencies: Sequence[Depends]) -> APIRouter:
@@ -376,14 +389,20 @@ def memory_add_router(dependencies: Sequence[Depends]) -> APIRouter:
             record = await asyncio.to_thread(additions.save, addition)
         except ValueError as error:
             raise HTTPException(409, str(error)) from error
-        saved = await asyncio.to_thread(additions.receipt, record.evidence_id)
+        saved = await asyncio.to_thread(
+            partial(additions.receipt, record.evidence_id, user_id=addition.user_id)
+        )
         background.add_task(schedule, additions, record)
         return saved
 
     @router.get("/additions/{request_id}", response_model=MemoryAddReceipt)
-    def memory_add_status(request_id: UUID, request: Request) -> MemoryAddReceipt:
+    def memory_add_status(
+        request_id: UUID,
+        request: Request,
+        user_id: str = Query(default=DEFAULT_USER_ID, pattern=USER_ID_PATTERN),
+    ) -> MemoryAddReceipt:
         try:
-            return store(request).receipt_for_request(request_id)
+            return store(request).receipt_for_request(request_id, user_id=user_id)
         except KeyError as error:
             raise HTTPException(404, "memory_add_not_found") from error
         except ValueError as error:

memhub/apps/core/src/one_memory/memory_evolution.py

@@ -0,0 +1,618 @@
+"""System 2 for ultra: evolve runs, their receipts, the nightly schedule, graph backfill.
+
+hy-memory 2.0.9's `client.evolve(user_id, agent_id, max_batches)` is ultra
+only, holds no lock, and has no scheduler: MemHub owns all three.
+
+- A run is asynchronous: `start` answers a receipt at once and the run goes
+  on in the background (evolve makes model calls across the user's whole
+  library and can take minutes, longer than any HTTP timeout). One run per
+  user at a time: asking again while one runs returns that run. A run holds
+  the write coordinator's per-user exclusivity, so it never overlaps that
+  user's write batch, and loops while the SDK answers `should_continue`, up
+  to `EVOLVE_MAX_ROUNDS`.
+- The schedule runs every user that has memories once a day at the
+  configured local time, and catches up after the machine slept through it.
+- Switching to ultra backfills each user's existing facts into the graph
+  once (`build_memory_graph`).
+
+What is kept on disk is counts and stable codes only, never memory text:
+`<data_dir>/evolve-state.json`.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import json
+import logging
+import os
+import re
+import threading
+from collections.abc import Awaitable, Callable, Iterable, Mapping
+from datetime import UTC, datetime, timedelta
+from pathlib import Path
+from tempfile import NamedTemporaryFile
+from typing import Any, Literal, Protocol
+from uuid import uuid4
+
+from pydantic import BaseModel, ConfigDict, Field
+
+from one_memory.diagnostic_log import diagnostic_event
+from one_memory.memory_server.config_state import EvolveSchedule
+from one_memory.memory_server.durable_queue import safe_error_code
+from one_memory.sdk_write_processor import PRODUCT_AGENT_ID
+
+logger = logging.getLogger(__name__)
+
+#: Rounds one run makes while the SDK says there is more to do.
+EVOLVE_MAX_ROUNDS = 20
+#: How often the schedule looks at the clock; also bounds catch-up after wake.
+SCHEDULE_CHECK_SECONDS = 300.0
+#: Receipts kept in memory for `GET /api/v2/evolve/{run_id}`.
+_KEPT_RUNS = 64
+_STABLE_CODE = re.compile(r"[a-z][a-z0-9_]{0,95}")
+_COUNTS = (
+    "remaining_clusters",
+    "remaining_actionable_clusters",
+    "remaining_fresh_facts",
+)
+
+RunState = Literal["running", "succeeded", "failed"]
+Trigger = Literal["manual", "scheduled"]
+
+
+class EvolveModeError(RuntimeError):
+    """Evolve asked for outside ultra."""
+
+    def __init__(self) -> None:
+        super().__init__("memory_mode_not_ultra")
+
+
+class EvolveRunReceipt(BaseModel):
+    """One evolve run. Counts and codes; `results` only while in memory."""
+
+    model_config = ConfigDict(extra="forbid")
+
+    run_id: str
+    user_id: str
+    agent_id: str
+    max_batches: int | None = None
+    trigger: Trigger = "manual"
+    state: RunState = "running"
+    started_at: datetime
+    finished_at: datetime | None = None
+    rounds: int = 0
+    tasks_processed: int = 0
+    batch_failures: int = 0
+    cycle_status: str | None = None
+    should_continue: bool | None = None
+    continuation_reason: str | None = None
+    remaining_clusters: int | None = None
+    remaining_actionable_clusters: int | None = None
+    remaining_fresh_facts: int | None = None
+    #: True when the round cap stopped a run the SDK would have continued.
+    capped: bool = False
+    elapsed_ms: float = 0.0
+    error_code: str | None = None
+    failure_kind: str | None = None
+    #: The last round's SDK `request_id`, for matching SDK logs.
+    request_id: str | None = None
+    #: The last round's SDK `results` and `cross_domain_sweeper`. Kept only
+    #: in this receipt, never written to disk or diagnostics.
+    results: Any = None
+    cross_domain_sweeper: Any = None
+
+    def summary(self) -> dict[str, object]:
+        """What is persisted and shown in status: no SDK result bodies."""
+        return self.model_dump(mode="json", exclude={"results", "cross_domain_sweeper"})
+
+
+class UserEvolveState(BaseModel):
+    model_config = ConfigDict(extra="forbid")
+
+    last_run_at: datetime | None = None
+    last_result: dict[str, object] | None = None
+    last_error: str | None = None
+
+
+class GraphBuildUser(BaseModel):
+    model_config = ConfigDict(extra="forbid")
+
+    state: Literal["pending", "running", "succeeded", "failed"] = "pending"
+    memories_scanned: int | None = None
+    nodes: int | None = None
+    error_code: str | None = None
+
+
+class GraphBuildState(BaseModel):
+    """The one-time backfill of existing facts into the graph after → ultra."""
+
+    model_config = ConfigDict(extra="forbid")
+
+    state: Literal["idle", "running", "succeeded", "failed"] = "idle"
+    started_at: datetime | None = None
+    finished_at: datetime | None = None
+    users: dict[str, GraphBuildUser] = Field(default_factory=dict)
+
+
+class EvolveStateDocument(BaseModel):
+    model_config = ConfigDict(extra="forbid")
+
+    schema_version: int = Field(default=1, ge=1, le=1)
+    users: dict[str, UserEvolveState] = Field(default_factory=dict)
+    #: When the schedule first ran in ultra: the "last run" of a user that
+    #: never ran, so enabling ultra at noon does not evolve everyone at once.
+    schedule_anchor: datetime | None = None
+    graph_build: GraphBuildState = Field(default_factory=GraphBuildState)
+
+
+class EvolveStateStore:
+    """The only writer of `<data_dir>/evolve-state.json`."""
+
+    def __init__(self, data_dir: Path) -> None:
+        self.path = data_dir / "evolve-state.json"
+        self._lock = threading.RLock()
+        self._document = self._read()
+
+    def document(self) -> EvolveStateDocument:
+        with self._lock:
+            return self._document.model_copy(deep=True)
+
+    def user(self, user_id: str) -> UserEvolveState:
+        with self._lock:
+            return self._document.users.get(user_id, UserEvolveState()).model_copy()
+
+    def record_run(self, receipt: EvolveRunReceipt) -> None:
+        with self._lock:
+            users = dict(self._document.users)
+            users[receipt.user_id] = UserEvolveState(
+                last_run_at=receipt.started_at,
+                last_result=receipt.summary(),
+                last_error=receipt.error_code,
+            )
+            self._save(self._document.model_copy(update={"users": users}))
+
+    def anchor(self, now: datetime) -> datetime:
+        """The schedule anchor, set to `now` the first time it is asked."""
+        with self._lock:
+            if self._document.schedule_anchor is None:
+                self._save(self._document.model_copy(update={"schedule_anchor": now}))
+            assert self._document.schedule_anchor is not None
+            return self._document.schedule_anchor
+
+    def graph_build(self) -> GraphBuildState:
+        with self._lock:
+            return self._document.graph_build.model_copy(deep=True)
+
+    def set_graph_build(self, state: GraphBuildState) -> None:
+        with self._lock:
+            self._save(self._document.model_copy(update={"graph_build": state}))
+
+    def _save(self, document: EvolveStateDocument) -> None:
+        payload = json.dumps(document.model_dump(mode="json"), indent=2, sort_keys=True) + "\n"
+        self.path.parent.mkdir(parents=True, exist_ok=True)
+        temporary: Path | None = None
+        try:
+            with NamedTemporaryFile(
+                mode="w",
+                encoding="utf-8",
+                dir=self.path.parent,
+                prefix=".evolve-state.",
+                suffix=".tmp",
+                delete=False,
+            ) as stream:
+                temporary = Path(stream.name)
+                stream.write(payload)
+                stream.flush()
+                os.fsync(stream.fileno())
+            os.chmod(temporary, 0o600)
+            os.replace(temporary, self.path)
+        except OSError:
+            if temporary is not None:
+                temporary.unlink(missing_ok=True)
+            # Status is best-effort: a run must not fail because its record did.
+            logger.warning("evolve_state_write_failed")
+        self._document = document
+
+    def _read(self) -> EvolveStateDocument:
+        try:
+            return EvolveStateDocument.model_validate_json(self.path.read_text(encoding="utf-8"))
+        except FileNotFoundError:
+            return EvolveStateDocument()
+        except (OSError, ValueError):
+            # A damaged status file only loses history; start a fresh one.
+            logger.warning("evolve_state_unreadable")
+            return EvolveStateDocument()
+
+
+class EvolveCoordinator(Protocol):
+    """The write coordinator surface evolve and the graph backfill use."""
+
+    def exclusive_for_evolve(self, user_id: str) -> Any: ...
+
+    def use_client(self) -> Any: ...
+
+
+class EvolveService:
+    """Start, serialize and record evolve runs."""
+
+    def __init__(
+        self,
+        *,
+        coordinator: EvolveCoordinator,
+        state: EvolveStateStore,
+        memory_mode: Callable[[], str],
+        max_rounds: int = EVOLVE_MAX_ROUNDS,
+        now: Callable[[], datetime] = lambda: datetime.now(UTC),
+    ) -> None:
+        self.coordinator = coordinator
+        self.state = state
+        self.memory_mode = memory_mode
+        self.max_rounds = max_rounds
+        self._now = now
+        self._runs: dict[str, EvolveRunReceipt] = {}
+        self._tasks: dict[str, asyncio.Task[None]] = {}
+        self._active: dict[str, str] = {}
+
+    def start(
+        self,
+        user_id: str,
+        *,
+        agent_id: str | None = None,
+        max_batches: int | None = None,
+        trigger: Trigger = "manual",
+    ) -> EvolveRunReceipt:
+        """Start a run, or return the one already running for this user."""
+        if self.memory_mode() != "ultra":
+            raise EvolveModeError()
+        running = self._active.get(user_id)
+        if running is not None:
+            return self._runs[running]
+        receipt = EvolveRunReceipt(
+            run_id=str(uuid4()),
+            user_id=user_id,
+            agent_id=agent_id or PRODUCT_AGENT_ID,
+            max_batches=max_batches,
+            trigger=trigger,
+            started_at=self._now(),
+        )
+        self._remember(receipt)
+        self._active[user_id] = receipt.run_id
+        self._tasks[receipt.run_id] = asyncio.create_task(
+            self._run(receipt.run_id), name="one-memory-evolve"
+        )
+        return receipt
+
+    def get(self, run_id: str) -> EvolveRunReceipt | None:
+        return self._runs.get(run_id)
+
+    def running(self) -> tuple[EvolveRunReceipt, ...]:
+        return tuple(self._runs[run_id] for run_id in self._active.values())
+
+    async def wait(self, run_id: str) -> EvolveRunReceipt | None:
+        task = self._tasks.get(run_id)
+        if task is not None:
+            await asyncio.shield(task)
+        return self._runs.get(run_id)
+
+    async def shutdown(self) -> None:
+        tasks = tuple(self._tasks.values())
+        for task in tasks:
+            task.cancel()
+        await asyncio.gather(*tasks, return_exceptions=True)
+        self._tasks.clear()
+
+    def _remember(self, receipt: EvolveRunReceipt) -> None:
+        self._runs[receipt.run_id] = receipt
+        while len(self._runs) > _KEPT_RUNS:
+            oldest = next(iter(self._runs))
+            if oldest in self._active.values():
+                break
+            del self._runs[oldest]
+
+    async def _run(self, run_id: str) -> None:
+        receipt = self._runs[run_id]
+        diagnostic_event(
+            level="info",
+            component="memory.evolve",
+            event="memory_evolve.started",
+            outcome="started",
+            details={"trigger": receipt.trigger, "max_batches": receipt.max_batches},
+        )
+        try:
+            async with (
+                self.coordinator.exclusive_for_evolve(receipt.user_id),
+                self.coordinator.use_client() as client,
+            ):
+                while receipt.rounds < self.max_rounds:
+                    result = await asyncio.to_thread(
+                        client.evolve,
+                        receipt.user_id,
+                        agent_id=receipt.agent_id,
+                        max_batches=receipt.max_batches,
+                    )
+                    receipt = _folded(receipt, result)
+                    self._runs[run_id] = receipt
+                    if receipt.state == "failed" or not receipt.should_continue:
+                        break
+                else:
+                    receipt = receipt.model_copy(update={"capped": True})
+            if receipt.state == "running":
+                receipt = receipt.model_copy(update={"state": "succeeded"})
+        except asyncio.CancelledError:
+            receipt = receipt.model_copy(
+                update={"state": "failed", "error_code": "memory_evolve_cancelled"}
+            )
+            raise
+        except Exception as error:  # noqa: BLE001 - recorded as a stable code
+            code = safe_error_code(error)
+            receipt = receipt.model_copy(
+                update={
+                    "state": "failed",
+                    "error_code": _stable(code, "hy_memory_evolve_failed"),
+                    "failure_kind": type(error).__name__,
+                }
+            )
+        finally:
+            receipt = receipt.model_copy(update={"finished_at": self._now()})
+            self._runs[run_id] = receipt
+            self._active.pop(receipt.user_id, None)
+            self._tasks.pop(run_id, None)
+            await asyncio.to_thread(self.state.record_run, receipt)
+            diagnostic_event(
+                level="info" if receipt.state == "succeeded" else "warning",
+                component="memory.evolve",
+                event="memory_evolve.finished",
+                outcome="succeeded" if receipt.state == "succeeded" else "failed",
+                details={
+                    "trigger": receipt.trigger,
+                    "rounds": receipt.rounds,
+                    "tasks_processed": receipt.tasks_processed,
+                    "batch_failures": receipt.batch_failures,
+                    "capped": receipt.capped,
+                    "elapsed_ms": round(receipt.elapsed_ms, 1),
+                    "error_code": receipt.error_code,
+                },
+            )
+
+
+def _folded(receipt: EvolveRunReceipt, result: object) -> EvolveRunReceipt:
+    """One more evolve round's result added to the run."""
+    if not isinstance(result, Mapping):
+        return receipt.model_copy(
+            update={"state": "failed", "error_code": "hy_memory_evolve_shape_unknown"}
+        )
+    failures = result.get("batch_failures")
+    update: dict[str, object] = {
+        "rounds": receipt.rounds + 1,
+        "tasks_processed": receipt.tasks_processed + _int(result.get("tasks_processed")),
+        "batch_failures": receipt.batch_failures
+        + (len(failures) if isinstance(failures, list | tuple) else _int(failures)),
+        "elapsed_ms": receipt.elapsed_ms + _float(result.get("elapsed_ms")),
+        "should_continue": result.get("should_continue") is True,
+        "cycle_status": _text(result.get("cycle_status")),
+        "continuation_reason": _text(result.get("continuation_reason")),
+        "request_id": _text(result.get("request_id")),
+        "results": result.get("results"),
+        "cross_domain_sweeper": result.get("cross_domain_sweeper"),
+    }
+    for name in _COUNTS:
+        value = result.get(name)
+        update[name] = value if isinstance(value, int) and not isinstance(value, bool) else None
+    if result.get("success") is not True:
+        code = result.get("error_code")
+        kind = result.get("failure_kind")
+        update.update(
+            state="failed",
+            should_continue=False,
+            error_code=(
+                code
+                if isinstance(code, str) and _STABLE_CODE.fullmatch(code)
+                else "hy_memory_evolve_rejected"
+            ),
+            failure_kind=kind if isinstance(kind, str) and _STABLE_CODE.fullmatch(kind) else None,
+        )
+    return receipt.model_copy(update=update)
+
+
+def _stable(code: str, fallback: str) -> str:
+    return code if _STABLE_CODE.fullmatch(code) else fallback
+
+
+def _int(value: object) -> int:
+    return value if isinstance(value, int) and not isinstance(value, bool) else 0
+
+
+def _float(value: object) -> float:
+    if isinstance(value, int | float) and not isinstance(value, bool):
+        return float(value)
+    return 0.0
+
+
+def _text(value: object) -> str | None:
+    return value[:128] if isinstance(value, str) else None
+
+
+def evolve_due(
+    now: datetime,
+    schedule: EvolveSchedule,
+    last_run_at: datetime | None,
+    anchor: datetime,
+) -> bool:
+    """Whether a user's nightly evolve is due at `now` (aware, local time).
+
+    Due once the day's slot (`schedule.time`, local) has passed and the last
+    run (or, for a user who never ran, the anchor) is older than that slot.
+    A last run more than 24 hours old is due at once as well: a machine asleep
+    at the slot catches up when it wakes, even before the next day's slot.
+    """
+    if not schedule.enabled:
+        return False
+    local = now.astimezone()
+    hour, minute = (int(part) for part in schedule.time.split(":"))
+    slot = local.replace(hour=hour, minute=minute, second=0, microsecond=0)
+    reference = last_run_at or anchor
+    if local >= slot and reference < slot:
+        return True
+    return last_run_at is not None and local - last_run_at > timedelta(hours=24)
+
+
+class EvolveScheduler:
+    """Run evolve nightly for every user that has memories (ultra only)."""
+
+    def __init__(
+        self,
+        *,
+        service: EvolveService,
+        schedule: Callable[[], EvolveSchedule],
+        users_with_memories: Callable[[], Awaitable[Iterable[str]]],
+        now: Callable[[], datetime] = lambda: datetime.now().astimezone(),
+        check_seconds: float = SCHEDULE_CHECK_SECONDS,
+    ) -> None:
+        self.service = service
+        self.schedule = schedule
+        self.users_with_memories = users_with_memories
+        self._now = now
+        self.check_seconds = check_seconds
+        self._task: asyncio.Task[None] | None = None
+
+    def start(self) -> None:
+        if self._task is None or self._task.done():
+            self._task = asyncio.create_task(self._loop(), name="one-memory-evolve-schedule")
+
+    async def tick(self) -> tuple[str, ...]:
+        """Run every due user once, one after another; the users run."""
+        schedule = self.schedule()
+        if not schedule.enabled or self.service.memory_mode() != "ultra":
+            return ()
+        now = self._now()
+        anchor = await asyncio.to_thread(self.service.state.anchor, now)
+        ran: list[str] = []
+        for user_id in await self.users_with_memories():
+            last = self.service.state.user(user_id).last_run_at
+            if not evolve_due(now, schedule, last, anchor):
+                continue
+            try:
+                receipt = self.service.start(user_id, trigger="scheduled")
+            except EvolveModeError:
+                break
+            await self.service.wait(receipt.run_id)
+            ran.append(user_id)
+        if ran:
+            diagnostic_event(
+                level="info",
+                component="memory.evolve",
+                event="memory_evolve.schedule_ran",
+                outcome="succeeded",
+                details={"users": len(ran)},
+            )
+        return tuple(ran)
+
+    async def _loop(self) -> None:
+        while True:
+            try:
+                await self.tick()
+            except asyncio.CancelledError:
+                raise
+            except Exception as error:  # noqa: BLE001 - the loop must survive
+                logger.warning("evolve_schedule_tick_failed error_type=%s", type(error).__name__)
+            await asyncio.sleep(self.check_seconds)
+
+    async def shutdown(self) -> None:
+        if self._task is None:
+            return
+        self._task.cancel()
+        await asyncio.gather(self._task, return_exceptions=True)
+        self._task = None
+
+
+class MemoryGraphBackfill:
+    """After a switch to ultra, project each user's existing facts into the graph once."""
+
+    def __init__(
+        self,
+        *,
+        coordinator: EvolveCoordinator,
+        state: EvolveStateStore,
+        now: Callable[[], datetime] = lambda: datetime.now(UTC),
+    ) -> None:
+        self.coordinator = coordinator
+        self.state = state
+        self._now = now
+        self._task: asyncio.Task[None] | None = None
+
+    @property
+    def running(self) -> bool:
+        return self._task is not None and not self._task.done()
+
+    def start(self, user_ids: Iterable[str]) -> GraphBuildState:
+        if self.running:
+            return self.state.graph_build()
+        users = tuple(dict.fromkeys(user_ids))
+        state = GraphBuildState(
+            state="running",
+            started_at=self._now(),
+            users={user_id: GraphBuildUser() for user_id in users},
+        )
+        self.state.set_graph_build(state)
+        self._task = asyncio.create_task(self._run(users), name="one-memory-graph-backfill")
+        return state
+
+    async def wait(self) -> None:
+        if self._task is not None:
+            await asyncio.shield(self._task)
+
+    async def _run(self, users: tuple[str, ...]) -> None:
+        state = self.state.graph_build()
+        for user_id in users:
+            state.users[user_id] = GraphBuildUser(state="running")
+            await asyncio.to_thread(self.state.set_graph_build, state.model_copy(deep=True))
+            try:
+                async with (
+                    self.coordinator.exclusive_for_evolve(user_id),
+                    self.coordinator.use_client() as client,
+                ):
+                    result = await asyncio.to_thread(
+                        client.build_memory_graph,
+                        user_id,
+                        agent_id=None,
+                        limit=10000,
+                        clear_first=False,
+                    )
+                ok = isinstance(result, Mapping) and result.get("success") is True
+                state.users[user_id] = GraphBuildUser(
+                    state="succeeded" if ok else "failed",
+                    memories_scanned=_int(result.get("memories_scanned"))
+                    if isinstance(result, Mapping)
+                    else None,
+                    nodes=_int(result.get("nodes")) if isinstance(result, Mapping) else None,
+                    error_code=None if ok else "hy_memory_graph_build_rejected",
+                )
+            except asyncio.CancelledError:
+                raise
+            except Exception as error:  # noqa: BLE001 - recorded as a stable code
+                code = safe_error_code(error)
+                state.users[user_id] = GraphBuildUser(
+                    state="failed",
+                    error_code=_stable(code, "hy_memory_graph_build_failed"),
+                )
+        failed = any(user.state == "failed" for user in state.users.values())
+        state.state = "failed" if failed else "succeeded"
+        state.finished_at = self._now()
+        await asyncio.to_thread(self.state.set_graph_build, state)
+        diagnostic_event(
+            level="warning" if failed else "info",
+            component="memory.evolve",
+            event="memory_graph_backfill.finished",
+            outcome="failed" if failed else "succeeded",
+            details={
+                "users": len(state.users),
+                "failed_users": sum(user.state == "failed" for user in state.users.values()),
+            },
+        )
+
+    async def shutdown(self) -> None:
+        if self._task is None:
+            return
+        self._task.cancel()
+        await asyncio.gather(self._task, return_exceptions=True)
+        self._task = None

memhub/apps/core/src/one_memory/memory_features.py

@@ -0,0 +1,262 @@
+"""SDK retrieval and extraction features a person may turn on or off.
+
+MemHub hard-codes none of these. Each is one `MemoryConfig` field that the SDK
+reads from its config object (never from `os.environ` at call time), so
+`hy_sdk_runtime.build_sdk_config` sets it per client and a client's features
+are fixed for its life, like its mode. An unset feature keeps the SDK default,
+so with nothing set the SDK config is byte for byte what it was before this
+setting existed.
+
+The setting is one global document in `config.json` (`memory_features`),
+beside `memory_mode`; management stays global, not per user. A change
+rebuilds the SDK handles the way a mode change does (#218 use count).
+
+What is deliberately left out, and why, is in
+docs/contracts/memory-features-v1.md. In short: flags the SDK reads straight
+from the process environment (the System 2 schema dedup/compaction lane, the
+balanced-catalog and specific-supplement lanes, the hybrid reader's
+`MEMORY_HYBRID_V2_*` knobs) cannot be scoped to one client; flags MemHub never
+exercises (query rewrite needs `queries=[...]`, session narratives need
+`finalize_session`); flags that are dangerous or internal (graph write, local
+cross-encoder download, access write-back); and everything MemHub already
+controls elsewhere (pipeline log, trace, memory operations, telemetry, mode).
+
+No SDK import here: only `hy_sdk_runtime` and `hy_sdk_migration` import it.
+`config_state` imports this module, so it imports nothing from there.
+"""
+
+from __future__ import annotations
+
+from collections.abc import Mapping
+from dataclasses import dataclass
+from typing import Literal
+
+from pydantic import BaseModel, ConfigDict, StrictBool
+
+MemoryFeatureName = Literal[
+    "graph_rerank",
+    "same_series_timeline",
+    "episode_recall",
+    "session_episodes",
+    "episode_summary",
+    "intention",
+    "few_shot",
+]
+#: Read features change search; write features change what extraction stores.
+FeatureSide = Literal["read", "write"]
+#: none: no extra call. extra_query: one more store query per search.
+#: llm_call: extra generation-model calls. prompt_tokens: no extra call, but
+#: a longer extraction prompt.
+FeatureCost = Literal["none", "extra_query", "llm_call", "prompt_tokens"]
+#: Why an enabled feature does nothing right now (`active: false`).
+FeatureReason = Literal["disabled", "requires_pro", "requires_ultra", "requires_graph_rerank"]
+#: A client's features as a hashable key: sorted (name, value) pairs, set ones only.
+type FeatureOverrides = tuple[tuple[str, bool], ...]
+
+_MODE_RANK: dict[str, int] = {"lite": 0, "pro": 1, "ultra": 2}
+
+
+@dataclass(frozen=True)
+class FeatureSpec:
+    name: MemoryFeatureName
+    #: `MemoryConfig` section and field the value is written to.
+    section: Literal["graph_store", "recall", "llm"]
+    field: str
+    #: The SDK's own env var for the same field; only its default comes from it.
+    env: str
+    side: FeatureSide
+    #: Lowest mode the feature acts in.
+    requires_mode: Literal["pro", "ultra"]
+    #: Another feature that must be active for this one to act.
+    requires_feature: MemoryFeatureName | None
+    cost: FeatureCost
+    description: str
+
+    @property
+    def sdk_field(self) -> str:
+        return f"{self.section}.{self.field}"
+
+
+FEATURES: tuple[FeatureSpec, ...] = (
+    FeatureSpec(
+        name="graph_rerank",
+        section="graph_store",
+        field="graph_rerank_enabled",
+        env="MEMORY_GRAPH_RERANK_ENABLED",
+        side="read",
+        requires_mode="ultra",
+        requires_feature=None,
+        cost="extra_query",
+        description=(
+            "Usage A: propagate support over knowledge-graph edges inside the "
+            "recall pool, so memories linked to strong hits rank higher. Needs "
+            "the graph store (ultra). No model call; one graph query per search."
+        ),
+    ),
+    FeatureSpec(
+        name="same_series_timeline",
+        section="recall",
+        field="same_series_timeline_enabled",
+        env="MEMORY_SAME_SERIES_TIMELINE_ENABLED",
+        side="read",
+        requires_mode="ultra",
+        requires_feature="graph_rerank",
+        cost="none",
+        description=(
+            "Keep memories joined by a same-series graph edge together, in time "
+            "order, in the results. Uses the edges graph rerank reads, so it "
+            "needs graph_rerank (ultra). No model call."
+        ),
+    ),
+    FeatureSpec(
+        name="episode_recall",
+        section="recall",
+        field="episode_recall_enabled",
+        env="MEMORY_EPISODE_RECALL_ENABLED",
+        side="read",
+        requires_mode="pro",
+        requires_feature=None,
+        cost="extra_query",
+        description=(
+            "Usage B: pull in memories from the same episode as a hit and score "
+            "them on tag-keyword overlap, content overlap and time affinity. No "
+            "graph needed (pro or ultra). No model call; one store query per "
+            "search. Helps only memories that carry an episode_id."
+        ),
+    ),
+    FeatureSpec(
+        name="session_episodes",
+        section="recall",
+        field="session_episode_enabled",
+        env="MEMORY_SESSION_EPISODE_ENABLED",
+        side="write",
+        requires_mode="pro",
+        requires_feature=None,
+        cost="none",
+        description=(
+            "Episode write side: memories extracted from one conversation share "
+            "one stable episode_id, instead of the default rule (a new episode "
+            "after 30 minutes idle or 50 memories). No model call."
+        ),
+    ),
+    FeatureSpec(
+        name="episode_summary",
+        section="recall",
+        field="episode_summary_enabled",
+        env="MEMORY_EPISODE_SUMMARY_ENABLED",
+        side="write",
+        requires_mode="pro",
+        requires_feature=None,
+        cost="llm_call",
+        description=(
+            "When an episode closes (the idle-gap rule), summarize it once into "
+            "an episode memory. One model call per closed episode of 3 or more "
+            "memories."
+        ),
+    ),
+    FeatureSpec(
+        name="intention",
+        section="llm",
+        field="enable_intention",
+        env="MEMORY_ENABLE_INTENTION",
+        side="write",
+        requires_mode="pro",
+        requires_feature=None,
+        cost="none",
+        description=(
+            "Extract the user's stated plans and intentions (L7) along with "
+            "facts. On by default; off drops them from the extraction prompt. "
+            "No extra call."
+        ),
+    ),
+    FeatureSpec(
+        name="few_shot",
+        section="llm",
+        field="few_shot_enabled",
+        env="MEMORY_FEW_SHOT_ENABLED",
+        side="write",
+        requires_mode="pro",
+        requires_feature=None,
+        cost="prompt_tokens",
+        description=(
+            "Append worked examples to the extraction and reconcile prompts. "
+            "Can help a weaker model; no extra call, but every extraction "
+            "sends more input tokens."
+        ),
+    ),
+)
+FEATURE_NAMES: tuple[MemoryFeatureName, ...] = tuple(spec.name for spec in FEATURES)
+FEATURES_BY_NAME: dict[str, FeatureSpec] = {spec.name: spec for spec in FEATURES}
+
+
+class MemoryFeatures(BaseModel):
+    """`config.json` `memory_features`, and the `PUT` body.
+
+    Every key is optional; absent (or null) means the SDK default. In a `PUT`
+    an absent key is left as it is and null resets it. Unknown keys and
+    non-boolean values are refused (422).
+    """
+
+    model_config = ConfigDict(frozen=True, extra="forbid", strict=True)
+
+    graph_rerank: StrictBool | None = None
+    same_series_timeline: StrictBool | None = None
+    episode_recall: StrictBool | None = None
+    session_episodes: StrictBool | None = None
+    episode_summary: StrictBool | None = None
+    intention: StrictBool | None = None
+    few_shot: StrictBool | None = None
+
+    def overrides(self) -> FeatureOverrides:
+        """The features a person set, as the client key."""
+        return feature_overrides(
+            {name: value for name, value in self.model_dump().items() if value is not None}
+        )
+
+    def merged(self, update: MemoryFeatures) -> MemoryFeatures:
+        """This document with the keys `update` names: a value sets, null resets."""
+        values = self.model_dump()
+        for name in update.model_fields_set:
+            values[name] = getattr(update, name)
+        return MemoryFeatures(**{key: value for key, value in values.items() if value is not None})
+
+
+def feature_overrides(values: Mapping[str, bool]) -> FeatureOverrides:
+    unknown = set(values) - set(FEATURES_BY_NAME)
+    if unknown:
+        raise ValueError("unknown_memory_feature")
+    return tuple(sorted((name, bool(value)) for name, value in values.items()))
+
+
+def effective_features(
+    overrides: FeatureOverrides, defaults: Mapping[str, bool]
+) -> dict[str, bool]:
+    """Every feature's value: what a person set, else the SDK default."""
+    values: dict[str, bool] = {spec.name: bool(defaults.get(spec.name, False)) for spec in FEATURES}
+    values.update(dict(overrides))
+    return values
+
+
+def feature_activity(name: str, effective: Mapping[str, bool], mode: str) -> FeatureReason | None:
+    """None when the feature acts in this mode; else why it does not."""
+    spec = FEATURES_BY_NAME[name]
+    if not effective.get(name, False):
+        return "disabled"
+    if _MODE_RANK.get(mode, 0) < _MODE_RANK[spec.requires_mode]:
+        return "requires_ultra" if spec.requires_mode == "ultra" else "requires_pro"
+    if spec.requires_feature is not None and (
+        feature_activity(spec.requires_feature, effective, mode) is not None
+    ):
+        return "requires_graph_rerank"
+    return None
+
+
+def non_default_features(
+    overrides: FeatureOverrides, defaults: Mapping[str, bool]
+) -> dict[str, bool]:
+    """The features whose value differs from the SDK default, for status."""
+    return {
+        name: value
+        for name, value in effective_features(overrides, defaults).items()
+        if value != bool(defaults.get(name, False))
+    }

memhub/apps/core/src/one_memory/memory_server/config_state.py

@@ -13,9 +13,11 @@ import threading
 from collections.abc import Mapping
 from pathlib import Path
 from tempfile import NamedTemporaryFile
+from typing import Literal
 
 from pydantic import BaseModel, ConfigDict, Field, StrictBool, field_validator
 
+from one_memory.memory_features import MemoryFeatures
 from one_memory.memory_queue import ExtractionWindowPreset
 
 _SOURCE_ID = re.compile(r"^[a-z0-9][a-z0-9-]{0,127}$")
@@ -27,6 +29,31 @@ class SourceConfig(BaseModel):
     sync_enabled: StrictBool = False
 
 
+#: The SDK pipeline mode (hy-memory `HyMemoryClient(mode=...)`):
+#: lite keeps raw conversations only, with no model call; pro extracts
+#: memories (System 1); ultra adds the knowledge graph and evolve (System 2).
+MemoryMode = Literal["lite", "pro", "ultra"]
+MEMORY_MODES: tuple[MemoryMode, ...] = ("lite", "pro", "ultra")
+_SCHEDULE_TIME = re.compile(r"^(?:[01][0-9]|2[0-3]):[0-5][0-9]$")
+
+
+class EvolveSchedule(BaseModel):
+    """When ultra runs evolve for every user: once a day, local time."""
+
+    model_config = ConfigDict(frozen=True, extra="forbid", strict=True)
+
+    enabled: StrictBool = True
+    #: `HH:MM`, 24-hour, the machine's local time.
+    time: str = "03:00"
+
+    @field_validator("time")
+    @classmethod
+    def valid_time(cls, value: str) -> str:
+        if _SCHEDULE_TIME.fullmatch(value) is None:
+            raise ValueError("invalid_evolve_schedule_time")
+        return value
+
+
 class ConfigDocument(BaseModel):
     model_config = ConfigDict(frozen=True, extra="forbid", strict=True)
 
@@ -39,6 +66,13 @@ class ConfigDocument(BaseModel):
     #: needs no re-entry of a key.
     sedimentation_enabled: StrictBool = True
     extraction_window_preset: ExtractionWindowPreset = ExtractionWindowPreset.THREE_MONTHS
+    #: The SDK mode every client is built with. `ONE_MEMORY_GRAPH_STORE` (dev)
+    #: still forces ultra over it; see `hy_graph_admin.effective_memory_mode`.
+    memory_mode: MemoryMode = "pro"
+    evolve_schedule: EvolveSchedule = Field(default_factory=EvolveSchedule)
+    #: SDK retrieval and extraction features a person set; an unset one keeps
+    #: the SDK default. See `one_memory.memory_features`.
+    memory_features: MemoryFeatures = Field(default_factory=MemoryFeatures)
 
     @field_validator("sources")
     @classmethod
@@ -61,6 +95,9 @@ class ConfigFileError(RuntimeError):
     pass
 
 
+_WRITTEN_WHEN_SET = frozenset({"memory_mode", "evolve_schedule", "memory_features"})
+
+
 class ConfigStore:
     """The only writer of ``<data_dir>/config.json`` inside Core."""
 
@@ -107,6 +144,45 @@ class ConfigStore:
             self._document = updated
             return updated
 
+    def memory_mode(self) -> MemoryMode:
+        with self._lock:
+            return self._document.memory_mode
+
+    def set_memory_mode(self, mode: MemoryMode) -> ConfigDocument:
+        if mode not in MEMORY_MODES:
+            raise ValueError("invalid_memory_mode")
+        with self._lock:
+            updated = self._document.model_copy(update={"memory_mode": mode})
+            self._write(updated)
+            self._document = updated
+            return updated
+
+    def evolve_schedule(self) -> EvolveSchedule:
+        with self._lock:
+            return self._document.evolve_schedule
+
+    def set_evolve_schedule(self, schedule: EvolveSchedule) -> ConfigDocument:
+        if not isinstance(schedule, EvolveSchedule):
+            raise ValueError("invalid_evolve_schedule")
+        with self._lock:
+            updated = self._document.model_copy(update={"evolve_schedule": schedule})
+            self._write(updated)
+            self._document = updated
+            return updated
+
+    def memory_features(self) -> MemoryFeatures:
+        with self._lock:
+            return self._document.memory_features
+
+    def set_memory_features(self, features: MemoryFeatures) -> ConfigDocument:
+        if not isinstance(features, MemoryFeatures):
+            raise ValueError("invalid_memory_features")
+        with self._lock:
+            updated = self._document.model_copy(update={"memory_features": features})
+            self._write(updated)
+            self._document = updated
+            return updated
+
     def migrate_extraction_window_preset(self, preset: ExtractionWindowPreset) -> ConfigDocument:
         """Adopt the legacy SQL value only when config.json has no value."""
         if not isinstance(preset, ExtractionWindowPreset):
@@ -152,9 +228,16 @@ class ConfigStore:
             raise ConfigFileError("config_invalid") from error
 
     def _write(self, document: ConfigDocument) -> None:
+        # Settings added after 1.0.5 are written only once a person chose
+        # them, so an untouched install's file stays what older builds read.
+        unset = {name for name in _WRITTEN_WHEN_SET if name not in document.model_fields_set}
+        values = document.model_dump(mode="json", exclude=unset)
+        if "memory_features" in values:
+            # Only the features a person set; an absent one is the SDK default.
+            values["memory_features"] = document.memory_features.model_dump(exclude_none=True)
         payload = (
             json.dumps(
-                document.model_dump(mode="json"),
+                values,
                 ensure_ascii=False,
                 indent=2,
                 sort_keys=True,

memhub/apps/core/src/one_memory/memory_server/contracts.py

@@ -75,7 +75,11 @@ def batch_id(*, internal_session_id: str, first_round: int, last_round: int) ->
 INTERNAL_SESSION_ID_MAX = 128
 
 
-def internal_session_id_for(conversation_id: str, source_id: str) -> str:
+def internal_session_id_for(
+    conversation_id: str,
+    source_id: str,
+    user_id: str = "local",
+) -> str:
     """Normalise a conversation identity to the durable ``internal_session_id``.
 
     This is the single derivation shared by every Evidence-first intake path.
@@ -86,7 +90,17 @@ def internal_session_id_for(conversation_id: str, source_id: str) -> str:
     per-session serialisation split). Only a value that the contract would have
     rejected before it ever reached the queue is replaced by the versioned
     session id, which is deterministic for a given (source, conversation).
+
+    Another user's conversation is always derived with the user in it: two
+    Hermes profiles can hold the same session id, and sharing a session key
+    would serialise and batch one person's turns behind the other's. `local`
+    keeps the identity it always had.
     """
+    if user_id != "local":
+        return session_id(
+            harness_id=source_id,
+            external_session_id=f"user:{user_id}|{conversation_id}",
+        )
     if len(conversation_id) <= INTERNAL_SESSION_ID_MAX:
         return conversation_id
     derived = session_id(harness_id=source_id, external_session_id=conversation_id)

memhub/apps/core/src/one_memory/memory_server/durable_queue.py

@@ -52,7 +52,9 @@ from one_memory.memory_tasks import (
 from one_memory.memory_tasks import (
     POLICY_STOPPED_CODES as _TASK_POLICY_STOPPED_CODES,
 )
+from one_memory.sdk_migration_log import error_chain_text
 from one_memory.service_levels import ServiceLevel, service_level_context
+from one_memory.user_scope import DEFAULT_USER_ID, lane_key, other_user_lane_prefix
 
 logger = logging.getLogger(__name__)
 
@@ -241,14 +243,16 @@ PERMISSION_REQUIRED_CODES = frozenset(
 )
 
 #: Model-side codes the SDK write coordinator raises before it can lease a
-#: client (no active profile, missing credential, secure storage down). They prove
-#: nothing was sent, so the write parks instead of being lost. Transport codes
-#: that can arrive after a request was in flight stay `unknown`.
+#: client (no active profile, missing credential, secure storage down, or the
+#: client failed to construct). They prove nothing was sent, so the write
+#: parks instead of being lost. Transport codes that can arrive after a
+#: request was in flight stay `unknown`.
 MODEL_NOT_READY_CODES = frozenset(
     {
         "model_profile_not_active",
         "model_secret_unavailable",
         "secure_storage_unavailable",
+        "hy_memory_client_unavailable",
     }
 )
 
@@ -292,6 +296,23 @@ _VALID_JOB_STATES = frozenset(
 #: A row in one of these states can never be moved back to a live one.
 _TERMINAL_WRITE_STATES = frozenset({COMPLETED, FAILED, DEAD, CANCELLED})
 
+#: The legacy task rows a write job can be linked to. The SDK cutover drops
+#: them (`target_schema.LEGACY_TABLES_IN_DROP_ORDER`, inputs before tasks), so
+#: the job-only paths must not assume any of them.
+_LEGACY_TASK_LINK_TABLES = ("memory_tasks", "memory_task_inputs", "memory_write_job_evidence")
+
+
+def _legacy_task_links_present(connection: sqlite3.Connection) -> bool:
+    present = {
+        str(row[0])
+        for row in connection.execute(
+            "SELECT name FROM sqlite_master WHERE type='table' AND name IN (?,?,?)",
+            _LEGACY_TASK_LINK_TABLES,
+        )
+    }
+    return present == set(_LEGACY_TASK_LINK_TABLES)
+
+
 #: Columns a current schema declares but an older `memory_write_jobs` may lack.
 #: `CREATE TABLE IF NOT EXISTS` is a no-op for an existing table, so without
 #: this an upgraded install would keep an older queue that every claim/lease
@@ -325,6 +346,27 @@ EVENT_TIME_WINDOW = timedelta(minutes=10)
 DIRECT_WRITE_SOURCES = frozenset({"one-memory-manual"})
 
 
+def request_user_id(request: WriteRequest) -> str:
+    """Whose write this is: its frozen input's user, `local` for id-only rows."""
+    users = {record.user_id for record in request.input_records}
+    if len(users) > 1:
+        raise ValueError("durable_write_multiple_users")
+    return next(iter(users), DEFAULT_USER_ID)
+
+
+def request_source_key(request: WriteRequest) -> str:
+    """The FIFO/claim key of a write: its source, per user.
+
+    `local` keeps the bare source id it always had. Another user's Hermes
+    lane is `hermes@user`, so one person's running job never holds the other
+    person's queue and the claim gate answers for that user's grant.
+    """
+    return lane_key(
+        request_user_id(request),
+        request.source_id or request.internal_session_id,
+    )
+
+
 @dataclass(frozen=True)
 class DurableWriteRecord:
     job_id: str
@@ -842,7 +884,7 @@ class DurableWriteQueue:
                     job_id,
                     queue_seq,
                     request.internal_session_id,
-                    request.source_id or request.internal_session_id,
+                    request_source_key(request),
                     request.batch_identity.batch_id,
                     key,
                     payload,
@@ -909,7 +951,9 @@ class DurableWriteQueue:
             f"{hashlib.sha256(file_key.encode()).hexdigest()}:"
             f"{hashlib.sha256(record.conversation_id.encode()).hexdigest()}"
         )
-        session_key = internal_session_id_for(record.conversation_id, record.source_id)
+        session_key = internal_session_id_for(
+            record.conversation_id, record.source_id, record.user_id
+        )
         written = connection.execute(
             "SELECT 1 FROM memory_write_jobs AS job, json_each(job.evidence_ids_json) AS ref "
             "WHERE job.session_key=? AND job.state!='assembling' AND ref.value=? LIMIT 1",
@@ -978,7 +1022,7 @@ class DurableWriteQueue:
                 str(uuid4()),
                 queue_seq,
                 session_key,
-                record.source_id,
+                lane_key(record.user_id, record.source_id),
                 assembly_key,
                 assembly_key,
                 payload_json,
@@ -1014,10 +1058,13 @@ class DurableWriteQueue:
         `flush` (history at end of file) completes the last turn and writes a
         short batch; without it (live polling) those stay in the row.
         """
+        # Two users' roots can hold a file with the same relative key (each
+        # Hermes profile has its own `state.db`): only this user's rows.
+        assembly_prefix = f"assembly:{user_id}:{source_id}:"
         rows = connection.execute(
             "SELECT * FROM memory_write_jobs WHERE state='assembling' "
-            "AND source_key=? ORDER BY queue_seq",
-            (source_id,),
+            "AND source_key=? AND substr(idempotency_key,1,?)=? ORDER BY queue_seq",
+            (lane_key(user_id, source_id), len(assembly_prefix), assembly_prefix),
         ).fetchall()
         finalized: list[str] = []
         for row in rows:
@@ -1140,7 +1187,7 @@ class DurableWriteQueue:
         Same records, same key: a re-read returns the job already written.
         """
         first = records[0]
-        session_id = internal_session_id_for(first.conversation_id, first.source_id)
+        session_id = internal_session_id_for(first.conversation_id, first.source_id, first.user_id)
         material = ":".join(f"{item.evidence_id}:{item.content_hash}" for item in records)
         request_id = f"turn:v2:{hashlib.sha256(material.encode()).hexdigest()}"
         existing = connection.execute(
@@ -1185,7 +1232,7 @@ class DurableWriteQueue:
                 job_id,
                 queue_seq,
                 session_id,
-                first.source_id,
+                lane_key(first.user_id, first.source_id),
                 request.batch_identity.batch_id,
                 request_id,
                 json.dumps(
@@ -1982,21 +2029,24 @@ class DurableWriteQueue:
                 or str(row["last_error"]) in NON_RETRYABLE_POST_WRITE_CODES
             ):
                 return 0
-            connection.execute(
-                """
-                UPDATE memory_tasks
-                SET status='pending', runner_token=NULL, attempts=0,
-                    last_error_code='model_content_manual_retry_requested', updated_at=?
-                WHERE status='unknown' AND task_id IN (
-                    SELECT inputs.task_id
-                    FROM memory_task_inputs AS inputs, json_each(inputs.refs_json) AS ref
-                    JOIN memory_write_job_evidence AS linked
-                      ON linked.evidence_id=json_extract(ref.value, '$.evidence_id')
-                    WHERE linked.job_id=?
+            # After the SDK cutover the legacy task tables are gone and the
+            # job is the only row to reset.
+            if not self.sdk_authority and _legacy_task_links_present(connection):
+                connection.execute(
+                    """
+                    UPDATE memory_tasks
+                    SET status='pending', runner_token=NULL, attempts=0,
+                        last_error_code='model_content_manual_retry_requested', updated_at=?
+                    WHERE status='unknown' AND task_id IN (
+                        SELECT inputs.task_id
+                        FROM memory_task_inputs AS inputs, json_each(inputs.refs_json) AS ref
+                        JOIN memory_write_job_evidence AS linked
+                          ON linked.evidence_id=json_extract(ref.value, '$.evidence_id')
+                        WHERE linked.job_id=?
+                    )
+                    """,
+                    (now, job_id),
                 )
-                """,
-                (now, job_id),
-            )
             changed = connection.execute(
                 """
                 UPDATE memory_write_jobs
@@ -2098,9 +2148,7 @@ class DurableWriteQueue:
                 f"WHERE state=? AND job_id IN ({placeholders})",
                 (PENDING, now, UNKNOWN, *job_ids),
             )
-            if connection.execute(
-                "SELECT 1 FROM sqlite_master WHERE type='table' AND name='memory_tasks'"
-            ).fetchone():
+            if _legacy_task_links_present(connection):
                 connection.execute(
                     f"UPDATE memory_tasks SET status=?, runner_token=NULL, "
                     f"last_error_code='model_content_reconciled_absent_retry', updated_at=? "
@@ -2158,9 +2206,7 @@ class DurableWriteQueue:
                 f"WHERE state=? AND attempts=1 AND job_id IN ({placeholders})",
                 (PENDING, now, UNKNOWN, *job_ids),
             )
-            if connection.execute(
-                "SELECT 1 FROM sqlite_master WHERE type='table' AND name='memory_tasks'"
-            ).fetchone():
+            if _legacy_task_links_present(connection):
                 connection.execute(
                     f"UPDATE memory_tasks SET status=?, runner_token=NULL, "
                     f"last_error_code='model_content_unknown_auto_retry', updated_at=? "
@@ -2177,6 +2223,14 @@ class DurableWriteQueue:
     def resolve_unknown_from_linked_output(self, job_id: str) -> bool:
         """Close an unknown durable row when local governance output proves success."""
         with main_database(self.database_path).read(row_factory=sqlite3.Row) as connection:
+            if not (
+                _legacy_task_links_present(connection)
+                and connection.execute(
+                    "SELECT 1 FROM sqlite_master WHERE type='table' AND name='memory_task_outputs'"
+                ).fetchone()
+            ):
+                # No legacy governance output survives the SDK cutover.
+                return False
             row = connection.execute(
                 """
                 SELECT 1
@@ -2467,6 +2521,12 @@ async def execute_job(
             source_id = request.internal_session_id
         if source_ids and source_ids != {source_id}:
             raise RuntimeError("durable_write_multiple_sources")
+        users = {getattr(record, "user_id", DEFAULT_USER_ID) for record in records}
+        if len(users) > 1:
+            raise RuntimeError("durable_write_multiple_users")
+        # The gates answer per user: another user's Hermes job asks about
+        # that user's grant, not `local`'s.
+        source_id = lane_key(next(iter(users), DEFAULT_USER_ID), source_id)
 
         async def mark() -> bool:
             return await in_the_background(queue.mark_invoked, job_id, owner=owner)
@@ -2503,6 +2563,8 @@ async def execute_job(
             in {
                 "model_content_source_disabled",
                 "model_sedimentation_disabled",
+                # Evolve holds this user; the job runs once it is done.
+                "memory_evolve_in_progress",
             }
             and not hy_invoked
         ):
@@ -2559,9 +2621,22 @@ async def execute_job(
         if code == "model_profile_not_active":
             logger.debug("durable_write_job_parked job_id=%s error_code=%s", job_id, code)
         elif state in PARKED_STATES:
-            logger.info("durable_write_job_parked job_id=%s error_code=%s", job_id, code)
+            logger.info(
+                "durable_write_job_parked job_id=%s error_code=%s chain=%s",
+                job_id,
+                code,
+                error_chain_text(error),
+            )
         else:
-            logger.error("durable_write_job_failed job_id=%s error_code=%s", job_id, code)
+            # Where it failed, as `file:line:function` frames of the exception
+            # chain, never its message: `error_code=AttributeError` alone
+            # named no location.
+            logger.error(
+                "durable_write_job_failed job_id=%s error_code=%s chain=%s",
+                job_id,
+                code,
+                error_chain_text(error),
+            )
         return
     await in_the_background(
         queue.settle,
@@ -2759,6 +2834,19 @@ class DurableWriteDrain:
                 source_clause = f"AND candidate.source_key IN ({placeholders}) "
                 source_params = tuple(sorted(enabled_sources))
             direct_sources = ",".join("?" for _ in DIRECT_WRITE_SOURCES)
+            # Another user's hand-added memories are `one-memory-manual@user`:
+            # the same lane ahead of history, still serial per user.
+            direct_lanes = " OR ".join(
+                "substr(candidate.source_key,1,?)=?" for _ in DIRECT_WRITE_SOURCES
+            )
+            direct_lane_params = tuple(
+                value
+                for source in sorted(DIRECT_WRITE_SOURCES)
+                for value in (
+                    len(other_user_lane_prefix(source)),
+                    other_user_lane_prefix(source),
+                )
+            )
             # Same eligibility as the window claim below (free lease, no
             # running job of its source, nothing earlier in its session), but
             # not bound to the window: a note must not wait for history.
@@ -2768,7 +2856,7 @@ class DurableWriteDrain:
                 "WHERE candidate.state=? "
                 "AND (candidate.lease_expires_at IS NULL OR candidate.lease_expires_at < ?) "
                 + source_clause
-                + f"AND candidate.source_key IN ({direct_sources}) "
+                + f"AND (candidate.source_key IN ({direct_sources}) OR {direct_lanes}) "
                 "AND NOT EXISTS ("
                 "SELECT 1 FROM memory_write_jobs AS held "
                 "WHERE held.source_key=candidate.source_key AND held.state=? "
@@ -2786,6 +2874,7 @@ class DurableWriteDrain:
                     now.isoformat(),
                     *source_params,
                     *sorted(DIRECT_WRITE_SOURCES),
+                    *direct_lane_params,
                     RUNNING,
                     now.isoformat(),
                     *_FIFO_BLOCKING_STATES,

memhub/apps/core/src/one_memory/memory_server/routes.py

@@ -19,10 +19,11 @@ does the writing. That separation is D52.
 
 from __future__ import annotations
 
+import hashlib
 from datetime import UTC, datetime
 from typing import Any, Literal, cast
 
-from fastapi import APIRouter, HTTPException, Request
+from fastapi import APIRouter, HTTPException, Query, Request
 from pydantic import BaseModel, ConfigDict, Field
 
 from one_memory.imports.models import (
@@ -43,6 +44,24 @@ from one_memory.memory_server.durable_queue import (
 )
 from one_memory.memory_server.registry import WriteRegistry, WriteState
 from one_memory.memory_server.submit import WriteConflictError
+from one_memory.user_scope import DEFAULT_USER_ID, USER_ID_PATTERN, UserId
+
+#: Durable keys of another user's writes start with this. A `local` request
+#: id never may, so `local` cannot name another user's receipt.
+_USER_KEY_PREFIX = "user:"
+
+
+def durable_request_key(request_id: str, user_id: str) -> str:
+    """The queue key of a caller's request id, namespaced per user.
+
+    Two users choosing the same request id must not collide: the key is the
+    queue's idempotency key and the SDK's `request_id`. `local` keeps its own
+    request id unchanged.
+    """
+    if user_id == DEFAULT_USER_ID:
+        return request_id
+    digest = hashlib.sha256(request_id.encode()).hexdigest()
+    return f"{_USER_KEY_PREFIX}{user_id}:{digest}"
 
 
 class MemoryWriteReceipt(BaseModel):
@@ -71,6 +90,7 @@ class SdkMemoryWriteInput(BaseModel):
     messages: list[OpenAIMessage] = Field(min_length=1, max_length=200)
     memory_at: datetime | None = None
     priority: Literal["live", "normal"] = "normal"
+    user_id: UserId = DEFAULT_USER_ID
 
 
 def _registry(
@@ -124,6 +144,9 @@ def sdk_router(dependencies: list[Any] | None = None) -> APIRouter:
     ) -> MemoryWriteReceipt:
         if not any(message.role == "user" for message in payload.messages):
             raise HTTPException(422, "messages_require_user_turn")
+        if payload.user_id == DEFAULT_USER_ID and payload.request_id.startswith(_USER_KEY_PREFIX):
+            raise HTTPException(422, "request_id_reserved_prefix")
+        request_key = durable_request_key(payload.request_id, payload.user_id)
         observed_at = payload.memory_at or datetime(1970, 1, 1, tzinfo=UTC)
         written_by = {
             "user": WrittenBy.PERSON,
@@ -133,6 +156,7 @@ def sdk_router(dependencies: list[Any] | None = None) -> APIRouter:
         records = tuple(
             materialize_write_input(
                 EvidenceInput(
+                    user_id=payload.user_id,
                     source_id=payload.source_id,
                     conversation_id=payload.session_id,
                     message_id=f"{payload.request_id}:{index}",
@@ -157,6 +181,7 @@ def sdk_router(dependencies: list[Any] | None = None) -> APIRouter:
         internal_session_id = internal_session_id_for(
             payload.session_id,
             payload.source_id,
+            payload.user_id,
         )
         contract = WriteRequest(
             internal_session_id=internal_session_id,
@@ -168,7 +193,7 @@ def sdk_router(dependencies: list[Any] | None = None) -> APIRouter:
             ),
             evidence_ids=[record.evidence_id for record in records],
             input_records=records,
-            idempotency_key=payload.request_id,
+            idempotency_key=request_key,
             priority=payload.priority,
             event_at=payload.memory_at,
             time_quality="exact" if payload.memory_at is not None else "approximate",
@@ -178,7 +203,7 @@ def sdk_router(dependencies: list[Any] | None = None) -> APIRouter:
         except (WriteConflictError, ValueError) as error:
             raise HTTPException(409, str(error)) from error
         return MemoryWriteReceipt(
-            request_id=record.request_id,
+            request_id=payload.request_id,
             state=record.state,
             created_at=record.created_at,
             updated_at=record.updated_at,
@@ -186,16 +211,24 @@ def sdk_router(dependencies: list[Any] | None = None) -> APIRouter:
         )
 
     @instance.get("/{request_id}", response_model=MemoryWriteReceipt)
-    def status(request_id: str, request: Request) -> MemoryWriteReceipt:
+    def status(
+        request_id: str,
+        request: Request,
+        user_id: str = Query(default=DEFAULT_USER_ID, pattern=USER_ID_PATTERN),
+    ) -> MemoryWriteReceipt:
+        # A receipt belongs to the user that submitted it. Another user's key
+        # is namespaced, and `local` may not name a namespaced key directly.
+        if user_id == DEFAULT_USER_ID and request_id.startswith(_USER_KEY_PREFIX):
+            raise HTTPException(404, "unknown request id")
         registry = _registry(request, write=False)
         try:
-            record = registry.get(request_id)
+            record = registry.get(durable_request_key(request_id, user_id))
         except DurableQueueSchemaUnavailableError as error:
             raise HTTPException(503, str(error)) from error
         if record is None:
             raise HTTPException(404, "unknown request id")
         return MemoryWriteReceipt(
-            request_id=record.request_id,
+            request_id=request_id,
             state=record.state,
             created_at=record.created_at,
             updated_at=record.updated_at,

memhub/apps/core/src/one_memory/memory_server/submit.py

@@ -271,7 +271,9 @@ def build_request_from_evidence(
         # Keep the exact conversation id when it fits the 128-char contract;
         # otherwise derive the versioned session id, so a long branch id does
         # not reject the whole history import.
-        internal_session_id=internal_session_id_for(seed.conversation_id, seed.source_id),
+        internal_session_id=internal_session_id_for(
+            seed.conversation_id, seed.source_id, seed.user_id
+        ),
         first_round=first_round,
         last_round=last_round,
         evidence_ids=list(ordered),

memhub/apps/core/src/one_memory/memory_server/unified_entry.py

@@ -28,6 +28,7 @@ from one_memory.memory_server.contracts import (
     internal_session_id_for,
 )
 from one_memory.memory_server.durable_queue import DurableWriteQueue, DurableWriteRecord
+from one_memory.user_scope import lane_key
 
 WritePriority = Literal["live", "normal"]
 
@@ -74,7 +75,7 @@ class UnifiedWriteIntake:
         if priority not in {"live", "normal"}:
             raise ValueError("priority must be 'live' or 'normal'")
         if self.source_writes_permitted is not None and not self.source_writes_permitted(
-            record.source_id
+            lane_key(record.user_id, record.source_id)
         ):
             # Freeze is the product's one Harness processing decision. Refuse
             # new intake so a frozen source creates no new durable write row;
@@ -106,7 +107,9 @@ class UnifiedWriteIntake:
         last_round: int,
         priority: WritePriority,
     ) -> WriteRequest:
-        session_id = internal_session_id_for(record.conversation_id, record.source_id)
+        session_id = internal_session_id_for(
+            record.conversation_id, record.source_id, record.user_id
+        )
         event_at, time_quality = _job_time((record,))
         request_key = (
             f"entry:{record.evidence_id}"
@@ -145,14 +148,14 @@ class UnifiedWriteIntake:
         ):
             raise ValueError("records must belong to one conversation")
         if self.source_writes_permitted is not None and not self.source_writes_permitted(
-            first.source_id
+            lane_key(first.user_id, first.source_id)
         ):
             raise SourceWriteDisabledError(first.evidence_id)
         material = ":".join(
             f"{record.evidence_id}:{record.revision}:{record.content_hash}" for record in records
         )
         request_id = f"turn:v1:{sha256(material.encode()).hexdigest()}"
-        session_id = internal_session_id_for(first.conversation_id, first.source_id)
+        session_id = internal_session_id_for(first.conversation_id, first.source_id, first.user_id)
         event_at, time_quality = _job_time(records)
         durable = self.queue.submit(
             WriteRequest(

memhub/apps/core/src/one_memory/memory_visibility.py

@@ -160,7 +160,7 @@ def pending(db: sqlite3.Connection, *, limit: int = 100) -> list[str]:
 
 
 def _trigger_sql() -> list[tuple[str, str]]:
-    """(table it depends on, statement) for each trigger."""
+    """(table it is on, statement) for each trigger."""
     return [
         (
             "evidence",
@@ -303,9 +303,61 @@ def install_triggers(db: sqlite3.Connection) -> None:
     for table, statement in _trigger_sql():
         if table not in present:
             continue
+        if not _body_tables(statement) <= present:
+            # SQLite resolves a trigger body only when it fires, so a body
+            # naming a dropped table fails the write it rides on (#233: the
+            # SDK cutover drops `governed_memory_evidence`, and every grant
+            # write then failed). Not installed, and an old copy is removed.
+            cursor.execute(f"DROP TRIGGER IF EXISTS {_trigger_name(statement)}")
+            continue
         cursor.execute(statement)
 
 
+def _trigger_name(statement: str) -> str:
+    return statement.split("IF NOT EXISTS", 1)[1].split()[0]
+
+
+def _body_tables(statement: str) -> set[str]:
+    """Tables the trigger body reads or writes besides the one it is on."""
+    _head, _begin, body = statement.partition("BEGIN")
+    words = body.replace("(", " ").replace(")", " ").split()
+    return {
+        name
+        for keyword, name in zip(words, words[1:], strict=False)
+        if keyword.upper() in {"FROM", "JOIN", "INTO"}
+    }
+
+
+def drop_stale_triggers(database_path: Path) -> tuple[str, ...]:
+    """Drop this module's triggers whose body names a table that is gone.
+
+    Startup on SDK authority does not install the cache (it is read only on
+    legacy authority), but a home migrated by an earlier build still carries
+    the triggers it installed. One whose body reads a dropped table fails
+    every write to the table it is on, so it is dropped here. Idempotent.
+    """
+    with main_database(database_path).write() as db, db:
+        cursor = db.cursor()
+        cursor.row_factory = sqlite3.Row
+        present = {
+            str(row["name"])
+            for row in cursor.execute("SELECT name FROM sqlite_master WHERE type='table'")
+        }
+        installed = {
+            str(row["name"])
+            for row in cursor.execute("SELECT name FROM sqlite_master WHERE type='trigger'")
+        }
+        dropped = tuple(
+            name
+            for _table, statement in _trigger_sql()
+            if (name := _trigger_name(statement)) in installed
+            and not _body_tables(statement) <= present
+        )
+        for name in dropped:
+            cursor.execute(f"DROP TRIGGER IF EXISTS {name}")
+    return dropped
+
+
 def startup(database_path: Path) -> None:
     """Idempotent: create the derived tables and triggers. No version change.
 

memhub/apps/core/src/one_memory/model_probe.py

@@ -564,7 +564,7 @@ def _embedding_request(
     secret: str | None,
 ) -> tuple[str, str, dict[str, str], dict[str, object]]:
     provider = profile.embedding_provider_kind or profile.provider_kind
-    base_url = (profile.embedding_base_url or profile.base_url or "").rstrip("/")
+    base_url = (profile.embedding_base_url or "").rstrip("/")
     if provider is ProviderKind.OLLAMA:
         return (
             "POST",

memhub/apps/core/src/one_memory/model_profiles.py

@@ -97,7 +97,6 @@ class ModelProfileInput(BaseModel):
         )
         embedding_provider, resolved_embedding_base_url = resolve_embedding_endpoint(
             provider_kind=self.provider_kind,
-            base_url=base_url,
             embedding_model=embedding_model,
             embedding_provider_kind=embedding_provider_kind,
             embedding_base_url=embedding_base_url,
@@ -107,6 +106,12 @@ class ModelProfileInput(BaseModel):
             ProviderKind.ANTHROPIC,
         }:
             raise ValueError("embedding provider must be OpenAI-compatible, Ollama, or local ONNX")
+        if (
+            embedding_provider is not ProviderKind.LOCAL_ONNX
+            and resolved_embedding_base_url is None
+        ):
+            # `ModelProfileStore.create` refuses it with its own code.
+            return self
         if (
             embedding_provider is not ProviderKind.LOCAL_ONNX
             and resolved_embedding_base_url is not None
@@ -308,11 +313,12 @@ class ModelProfileStore:
                     f"""
                     UPDATE {table}
                     SET embedding_provider_kind = provider_kind,
-                        embedding_base_url = base_url
+                        embedding_base_url = COALESCE(embedding_base_url, base_url)
                     WHERE embedding_model IS NOT NULL
                       AND embedding_provider_kind IS NULL
                     """
                 )
+                _pin_embedding_endpoints(connection, table)
             connection.execute(
                 """
                 INSERT OR IGNORE INTO model_profile_versions (
@@ -355,11 +361,11 @@ class ModelProfileStore:
         )
         embedding_provider_kind, embedding_base_url = resolve_embedding_endpoint(
             provider_kind=value.provider_kind,
-            base_url=base_url,
             embedding_model=defaulted_model,
             embedding_provider_kind=defaulted_provider,
             embedding_base_url=defaulted_base_url,
         )
+        require_embedding_base_url(embedding_provider_kind, embedding_base_url)
         try:
             with self._connect(writing=True) as connection:
                 connection.execute(
@@ -612,7 +618,6 @@ class ModelProfileStore:
         )
         embedding_provider_kind, embedding_base_url = resolve_embedding_endpoint(
             provider_kind=current.provider_kind,
-            base_url=base_url,
             embedding_model=embedding_model,
             embedding_provider_kind=(
                 patch.embedding_provider_kind
@@ -627,6 +632,7 @@ class ModelProfileStore:
         )
         if embedding_provider_kind in {ProviderKind.HY_DEFAULT, ProviderKind.ANTHROPIC}:
             raise ValueError("embedding provider must be OpenAI-compatible, Ollama, or local ONNX")
+        require_embedding_base_url(embedding_provider_kind, embedding_base_url)
         if (
             embedding_provider_kind is not ProviderKind.LOCAL_ONNX
             and embedding_base_url is not None
@@ -1167,21 +1173,77 @@ def attach_disable_thinking(
 def resolve_embedding_endpoint(
     *,
     provider_kind: ProviderKind,
-    base_url: str | None,
     embedding_model: str | None,
     embedding_provider_kind: ProviderKind | None,
     embedding_base_url: str | None,
 ) -> tuple[ProviderKind | None, str | None]:
+    """The embedding provider and the endpoint exactly as given.
+
+    The endpoint never follows the generation `base_url`: the collection name
+    hashes it, so a proxy or region edit of the generation endpoint moved
+    writes to a new, empty collection, and providers without embeddings were
+    sent embedding requests. `require_embedding_base_url` refuses a remote
+    embedding without one.
+    """
     if embedding_model is None:
         return None, None
     if embedding_provider_kind is ProviderKind.LOCAL_ONNX:
         return ProviderKind.LOCAL_ONNX, None
-    if embedding_provider_kind is None:
-        return provider_kind, embedding_base_url or base_url
-    return (
-        embedding_provider_kind,
-        embedding_base_url or DEFAULT_BASE_URLS.get(embedding_provider_kind),
-    )
+    return embedding_provider_kind or provider_kind, embedding_base_url
+
+
+class EmbeddingBaseUrlRequiredError(ValueError):
+    def __init__(self) -> None:
+        super().__init__("embedding_base_url_required")
+
+
+def require_embedding_base_url(provider: ProviderKind | None, base_url: str | None) -> None:
+    """A remote embedding must name its endpoint; the UI may offer the generation one."""
+    if provider not in {None, ProviderKind.LOCAL_ONNX} and not base_url:
+        raise EmbeddingBaseUrlRequiredError()
+
+
+def pinned_embedding_base_url(
+    *,
+    provider_kind: ProviderKind | str,
+    base_url: str | None,
+    embedding_model: str | None,
+    embedding_provider_kind: ProviderKind | str | None,
+    embedding_base_url: str | None,
+) -> str | None:
+    """The endpoint a stored profile's embedding resolved to before it was pinned.
+
+    The Runtime used `embedding_base_url or base_url`, and the provider default
+    only when both were empty (no collection could have existed then). Writing
+    that value keeps the endpoint, and so the collection name, byte-identical.
+    """
+    provider = ProviderKind(embedding_provider_kind or provider_kind)
+    if embedding_model is None or embedding_base_url or provider is ProviderKind.LOCAL_ONNX:
+        return embedding_base_url
+    return base_url or DEFAULT_BASE_URLS.get(provider)
+
+
+def _pin_embedding_endpoints(connection: sqlite3.Connection, table: str) -> None:
+    """Write the resolved endpoint into every remote embedding left without one."""
+    rows = connection.execute(
+        f"SELECT rowid, provider_kind, base_url, embedding_model, embedding_provider_kind "
+        f"FROM {table} WHERE embedding_model IS NOT NULL AND embedding_base_url IS NULL "
+        "AND COALESCE(embedding_provider_kind, provider_kind) != 'local_onnx'"
+    ).fetchall()
+    for row in rows:
+        connection.execute(
+            f"UPDATE {table} SET embedding_base_url = ? WHERE rowid = ?",
+            (
+                pinned_embedding_base_url(
+                    provider_kind=row["provider_kind"],
+                    base_url=row["base_url"],
+                    embedding_model=row["embedding_model"],
+                    embedding_provider_kind=row["embedding_provider_kind"],
+                    embedding_base_url=None,
+                ),
+                row["rowid"],
+            ),
+        )
 
 
 def validate_model_base_url(value: str) -> None:

memhub/apps/core/src/one_memory/recall.py

@@ -14,6 +14,7 @@ from pydantic import BaseModel, ConfigDict, Field
 
 from one_memory.hy_sdk_types import HySdkMemoryRecord
 from one_memory.memory_governance import GovernedMemory, MemoryStatus
+from one_memory.sdk_search_options import SdkSearchOptions
 from one_memory.semantic_deadline import (
     SemanticMemoryHit,
     SemanticSearch,
@@ -22,6 +23,7 @@ from one_memory.semantic_deadline import (
     normalize_semantic_hits,
     search_within,
 )
+from one_memory.user_scope import DEFAULT_USER_ID, UserId
 
 
 def recall_blend_mode() -> str:
@@ -49,7 +51,15 @@ class TruncationBudget(StrEnum):
     TIMEOUT = "timeout_budget"
 
 
-class RecallRequest(BaseModel):
+class RecallRequest(SdkSearchOptions):
+    """`POST /api/v1/recall`.
+
+    The `SdkSearchOptions` fields, `schema_limit` and `intention_limit` pass
+    through to the SDK memory search (and the floor and row filters to the
+    raw fallback). Absent, recall makes the search it always made: the
+    relevance floor, schema and intention lanes sized to `limit`.
+    """
+
     model_config = ConfigDict(frozen=True, extra="forbid")
 
     query: str = Field(min_length=1, max_length=2_000)
@@ -67,6 +77,19 @@ class RecallRequest(BaseModel):
     #: Wall-clock budget for candidate gather + score. Not a substitute for
     #: raising the old newest-500 page; it is the third parallel budget.
     timeout_ms: int = Field(default=2_000, ge=50, le=30_000)
+    #: Whose memories to recall. Omitted, it is `local`, the App's library.
+    user_id: UserId = DEFAULT_USER_ID
+    #: The SDK's schema (L6) and proactive intention (L7) lanes.
+    schema_limit: int | None = Field(default=None, ge=0, le=20)
+    intention_limit: int | None = Field(default=None, ge=0, le=20)
+
+    def search_overrides(self) -> Mapping[str, object]:
+        values = dict(super().search_overrides())
+        if self.schema_limit is not None:
+            values["schema_limit"] = self.schema_limit
+        if self.intention_limit is not None:
+            values["intention_limit"] = self.intention_limit
+        return values
 
 
 class RecallHit(BaseModel):

memhub/apps/core/src/one_memory/sdk_memory_qa.py

@@ -6,7 +6,7 @@ import asyncio
 import json
 from collections.abc import AsyncIterator, Awaitable, Callable
 from contextlib import aclosing
-from typing import Any, Literal, Protocol
+from typing import Any, Literal, Protocol, cast
 
 import httpx
 from pydantic import BaseModel, ConfigDict
@@ -14,7 +14,7 @@ from pydantic import BaseModel, ConfigDict
 from one_memory.http_client import create_async_client
 from one_memory.hy_runtime_config import HY_MEMORY_READER
 from one_memory.hy_sdk_runtime import sdk_relevance_floor
-from one_memory.hy_sdk_types import HySdkMemoryRecord, parse_sdk_search
+from one_memory.hy_sdk_types import HySdkMemoryRecord, parse_sdk_lane_search, parse_sdk_search
 from one_memory.memory_qa import (
     MemoryAnswerError,
     _DeferredSecretStore,
@@ -31,7 +31,9 @@ from one_memory.model_profiles import (
     ProviderKind,
     attach_disable_thinking,
 )
+from one_memory.sdk_search_options import SdkSearchFilters, SearchLane
 from one_memory.secrets import SecretStore, SecretStoreUnavailableError
+from one_memory.user_scope import owns
 
 _NO_GROUNDING = "现有记忆不足以回答这个问题。"
 _GROUNDING_LIMIT = 6
@@ -95,29 +97,51 @@ class SdkMemoryAnswerer:
         question: str,
         *,
         user_id: str = "local",
+        filters: SdkSearchFilters | None = None,
     ) -> tuple[SdkAskMemory, ...]:
+        """The user's active memories nearest the question.
+
+        `filters` are a caller's pass-through search options; absent, the
+        search is the one ask always made.
+        """
         client = await self.client_handle.client()
+        overrides = filters.search_overrides() if filters is not None else {}
         raw = await asyncio.to_thread(
             client.search,
             question,
-            user_ids=[user_id],
-            agent_ids=[],
-            limit=12,
-            min_score=sdk_relevance_floor(),
-            schema_min_score=sdk_relevance_floor(),
-            schema_limit=12,
-            intention_limit=12,
-            scene="normal",
-            reader=HY_MEMORY_READER,
+            **{
+                "user_ids": [user_id],
+                "agent_ids": [],
+                "limit": 12,
+                "min_score": sdk_relevance_floor(),
+                "schema_min_score": sdk_relevance_floor(),
+                "schema_limit": 12,
+                "intention_limit": 12,
+                "scene": "normal",
+                "reader": HY_MEMORY_READER,
+                **overrides,
+            },
         )
-        hits = parse_sdk_search(raw, error_code="hy_memory_search_shape_unknown")
+        if filters is not None and overrides:
+            hits = tuple(
+                hit
+                for hit in parse_sdk_lane_search(raw, error_code="hy_memory_search_shape_unknown")
+                if filters.passes_floor(cast(SearchLane, hit.lane), hit.score)
+            )
+        else:
+            hits = parse_sdk_search(raw, error_code="hy_memory_search_shape_unknown")
         found: list[SdkAskMemory] = []
         seen: set[str] = set()
         for hit in hits:
             if hit.memory_id in seen:
                 continue
             record = await _get_memory(client, hit.memory_id)
-            if record is None or record.status != "active" or not record.is_latest:
+            if (
+                record is None
+                or record.status != "active"
+                or not record.is_latest
+                or not owns(record.user_id, user_id)
+            ):
                 continue
             seen.add(record.memory_id)
             found.append(_wire_memory(record))
@@ -128,8 +152,9 @@ class SdkMemoryAnswerer:
         question: str,
         *,
         user_id: str = "local",
+        filters: SdkSearchFilters | None = None,
     ) -> SdkMemoryAnswer:
-        candidates = await self.retrieve(question, user_id=user_id)
+        candidates = await self.retrieve(question, user_id=user_id, filters=filters)
         if not candidates:
             return SdkMemoryAnswer(answer=_NO_GROUNDING, memories=(), mode="empty")
         profile = await self._active_profile()
@@ -160,8 +185,9 @@ class SdkMemoryAnswerer:
         question: str,
         *,
         user_id: str = "local",
+        filters: SdkSearchFilters | None = None,
     ) -> AsyncIterator[dict[str, object]]:
-        candidates = await self.retrieve(question, user_id=user_id)
+        candidates = await self.retrieve(question, user_id=user_id, filters=filters)
         await self._require_current(candidates)
         yield {"type": "recall", "memories": candidates}
         if not candidates:

memhub/apps/core/src/one_memory/sdk_memory_runtime.py

@@ -5,7 +5,8 @@ from __future__ import annotations
 import asyncio
 import logging
 import re
-from contextlib import suppress
+from collections.abc import AsyncIterator, Callable
+from contextlib import AsyncExitStack, asynccontextmanager
 from datetime import UTC, datetime, timedelta
 from pathlib import Path
 from typing import Any
@@ -32,16 +33,6 @@ from one_memory.semantic_deadline import SemanticMemoryHit
 _STABLE_CODE = re.compile(r"[a-z][a-z0-9_]{0,95}")
 
 logger = logging.getLogger(__name__)
-RAW_RETENTION_DAYS = 3
-RAW_RETENTION_CHECK_SECONDS = 6 * 60 * 60
-_RAW_STATUSES = (
-    "active",
-    "superseded",
-    "shadow",
-    "negated",
-    "conflicted",
-    "archived",
-)
 
 
 class SdkMemoryRuntime:
@@ -55,74 +46,17 @@ class SdkMemoryRuntime:
         write_queue: DurableWriteQueue,
         extraction_settings: ExtractionSettingsStore,
         data_dir: Path,
+        memory_mode: Callable[[], str] = lambda: "pro",
     ) -> None:
         self.coordinator = coordinator
+        #: The SDK mode, for the graph store's path (`hy_graph_admin`).
+        self.memory_mode = memory_mode
         self.read_handle = read_handle
         self.write_queue = write_queue
         self.extraction_settings = extraction_settings
         self.data_dir = data_dir
-        self._raw_retention_task: asyncio.Task[None] | None = None
-
-    def start_raw_retention(self) -> None:
-        if self._raw_retention_task is None or self._raw_retention_task.done():
-            self._raw_retention_task = asyncio.create_task(
-                self._raw_retention_loop(),
-                name="one-memory-sdk-raw-retention",
-            )
-
-    async def prune_expired_raw(self, *, now: datetime | None = None) -> int:
-        cutoff = (now or datetime.now(UTC)) - timedelta(days=RAW_RETENTION_DAYS)
-        cutoff_timestamp = int(cutoff.timestamp())
-        client = await self.read_handle.client()
-        deleted = 0
-        while True:
-            page = await asyncio.to_thread(
-                client.list_memories,
-                user_id="local",
-                layers=["l1_raw"],
-                statuses=list(_RAW_STATUSES),
-                order="asc",
-                limit=500,
-                offset=0,
-            )
-            bucket = page.get("vdb") if isinstance(page, dict) else None
-            rows = bucket.get("memories") if isinstance(bucket, dict) else None
-            if not isinstance(rows, list) or not rows:
-                return deleted
-            expired = [
-                row
-                for row in rows
-                if isinstance(row, dict)
-                and isinstance(row.get("memory_id"), str)
-                and int(row.get("memory_at") or row.get("gmt_created") or 0)
-                < cutoff_timestamp
-            ]
-            if not expired:
-                return deleted
-            removed_page = 0
-            for row in expired:
-                result = await asyncio.to_thread(client.delete, row["memory_id"])
-                if isinstance(result, dict) and result.get("success") is True:
-                    count = int(result.get("deleted_count") or 0)
-                    deleted += count
-                    removed_page += count
-            if removed_page == 0:
-                return deleted
-
-    async def _raw_retention_loop(self) -> None:
-        while True:
-            await asyncio.sleep(RAW_RETENTION_CHECK_SECONDS)
-            try:
-                removed = await self.prune_expired_raw()
-                if removed:
-                    logger.info("sdk_raw_retention_deleted count=%d", removed)
-            except asyncio.CancelledError:
-                raise
-            except Exception as error:
-                logger.warning(
-                    "sdk_raw_retention_failed error_type=%s",
-                    type(error).__name__,
-                )
+        # MemHub keeps every raw the SDK stores, in every mode: there is no
+        # retention window (owner decision; the SDK itself never deletes raw).
 
     async def semantic_recall(
         self,
@@ -153,55 +87,55 @@ class SdkMemoryRuntime:
         limit: int = 12,
     ) -> tuple[tuple[str, float], ...]:
         return tuple(
-            (hit.memory_id, hit.score)
-            for hit in await self.semantic_recall(query, limit=limit)
+            (hit.memory_id, hit.score) for hit in await self.semantic_recall(query, limit=limit)
         )
 
-    async def graph_snapshot(self, *, limit: int = 300) -> HyGraphSnapshot:
+    async def graph_snapshot(self, *, limit: int = 300, user_id: str = "local") -> HyGraphSnapshot:
         try:
             await self.coordinator.active_profile_credentials()
         except RuntimeError as error:
             if str(error) == "model_profile_not_active":
                 return HyGraphSnapshot(consulted=False)
             raise
-        try:
-            client = await self.coordinator.client()
-        except Exception as error:
-            raise HySdkError("hy_memory_graph_read_failed") from error
-        return await asyncio.to_thread(
-            read_graph_snapshot,
-            client,
-            "local",
-            "*",
-            limit,
-        )
+        async with AsyncExitStack() as stack:
+            try:
+                client = await stack.enter_async_context(self.coordinator.use_client())
+            except Exception as error:
+                raise HySdkError("hy_memory_graph_read_failed") from error
+            return await asyncio.to_thread(
+                read_graph_snapshot,
+                client,
+                user_id,
+                "*",
+                limit,
+            )
 
     async def digest(self) -> dict[str, object]:
-        client = await self.coordinator.client()
-        try:
-            raw = await asyncio.to_thread(
-                client.list_memories,
-                user_id="local",
-                limit=1_000,
-                offset=0,
-            )
-            agent_ids = _agent_ids(raw)
-            digests = [
-                {
-                    "agent_id": agent_id,
-                    "schemas_recorded": 0,
-                    **dict(
-                        await asyncio.to_thread(
-                            client.evolve,
-                            user_id="local",
-                            agent_id=agent_id,
-                        )
-                    ),
-                }
-                for agent_id in agent_ids
-            ]
-        except Exception as error:
-            raise HySdkError("hy_memory_digest_failed") from error
+        async with self.coordinator.use_client() as client:
+            try:
+                raw = await asyncio.to_thread(
+                    client.list_memories,
+                    user_id="local",
+                    limit=1_000,
+                    offset=0,
+                )
+                agent_ids = _agent_ids(raw)
+                digests = [
+                    {
+                        "agent_id": agent_id,
+                        "schemas_recorded": 0,
+                        **dict(
+                            await asyncio.to_thread(
+                                client.evolve,
+                                user_id="local",
+                                agent_id=agent_id,
+                            )
+                        ),
+                    }
+                    for agent_id in agent_ids
+                ]
+            except Exception as error:
+                raise HySdkError("hy_memory_digest_failed") from error
         return {
             "success": all(result.get("success") is not False for result in digests),
             "scope_count": len(digests),
@@ -222,6 +156,7 @@ class SdkMemoryRuntime:
             profile,
             api_key=secret,
             data_dir=self.data_dir,
+            mode=self.memory_mode(),
         )
 
     async def decide_graph_width(
@@ -238,38 +173,42 @@ class SdkMemoryRuntime:
             api_key=secret,
             data_dir=self.data_dir,
             decision=decision,
+            mode=self.memory_mode(),
         )
 
-    async def _mutation_client(self) -> Any:
+    @asynccontextmanager
+    async def _mutation_client(self) -> AsyncIterator[Any]:
         """The write client, with no-model and keychain refusals as SDK codes.
 
         Callers record these against the edit or deletion instead of failing
         the request that already committed it.
         """
-        try:
-            return await self.coordinator.client()
-        except RuntimeError as error:
-            code = str(error)
-            if not _STABLE_CODE.fullmatch(code):
-                code = "hy_memory_client_unavailable"
-            raise HySdkError(code) from error
+        async with AsyncExitStack() as stack:
+            try:
+                client = await stack.enter_async_context(self.coordinator.use_client())
+            except RuntimeError as error:
+                code = str(error)
+                if not _STABLE_CODE.fullmatch(code):
+                    code = "hy_memory_client_unavailable"
+                raise HySdkError(code) from error
+            yield client
 
     async def update_hy_memory(self, memory_id: str, content: str) -> dict[str, object]:
-        client = await self._mutation_client()
-        try:
-            result = await asyncio.to_thread(client.update, memory_id, content)
-        except Exception as error:
-            raise HySdkError("hy_memory_update_failed") from error
+        async with self._mutation_client() as client:
+            try:
+                result = await asyncio.to_thread(client.update, memory_id, content)
+            except Exception as error:
+                raise HySdkError("hy_memory_update_failed") from error
         if not isinstance(result, dict) or result.get("success") is not True:
             raise HySdkError("hy_memory_update_rejected")
         return result
 
     async def finish_deletion(self, memory_id: str) -> None:
-        client = await self._mutation_client()
-        try:
-            result = await asyncio.to_thread(client.delete, memory_id)
-        except Exception as error:
-            raise HySdkError("hy_memory_delete_failed") from error
+        async with self._mutation_client() as client:
+            try:
+                result = await asyncio.to_thread(client.delete, memory_id)
+            except Exception as error:
+                raise HySdkError("hy_memory_delete_failed") from error
         if not isinstance(result, dict):
             raise HySdkError("hy_memory_delete_rejected")
         error_code = result.get("error_code")
@@ -373,12 +312,7 @@ class SdkMemoryRuntime:
         return None
 
     async def shutdown(self) -> None:
-        if self._raw_retention_task is None:
-            return
-        self._raw_retention_task.cancel()
-        with suppress(asyncio.CancelledError):
-            await self._raw_retention_task
-        self._raw_retention_task = None
+        return None
 
 
 def _agent_ids(raw: object) -> tuple[str, ...]:

memhub/apps/core/src/one_memory/sdk_migration_runner.py

@@ -23,6 +23,7 @@ from pathlib import Path
 from typing import Any
 
 from one_memory.database_access import main_database
+from one_memory.hy_sdk_runtime import builtin_retrieval_snapshot
 from one_memory.imports.models import EvidenceRecord
 from one_memory.imports.storage import EvidenceStore
 from one_memory.memory_server.contracts import (
@@ -35,6 +36,8 @@ from one_memory.memory_server.durable_queue import (
     COMPLETED,
     FAILED,
     REAPED_PAYLOAD_MARKER,
+    RUNNING,
+    UNKNOWN,
     DurableWriteQueue,
 )
 from one_memory.memory_server.unified_entry import UnifiedWriteIntake
@@ -49,6 +52,7 @@ from one_memory.sdk_migration_bundle import inspect_sdk_migration_bundle
 from one_memory.sdk_migration_verify import (
     MemoryReadbackFailed,
     ReadbackReport,
+    expected_memory_ids,
     legacy_memory_ids_from,
     legacy_memory_ids_from_bundle,
     read_active_profile,
@@ -59,6 +63,8 @@ from one_memory.sdk_vector_reattach import (
     ReattachedCollection,
     legacy_vector_collections,
     reattach_legacy_vectors,
+    stored_memory_ids,
+    vectors_belong_to_another_space,
 )
 from one_memory.target_schema import drop_legacy_tables, legacy_tables_present
 
@@ -257,6 +263,7 @@ def prepare_legacy_write_queue_for_cutover(database_path: Path) -> tuple[int, in
         internal_session_id = internal_session_id_for(
             source.conversation_id,
             source.source_id,
+            source.user_id,
         )
         queue.submit(
             WriteRequest(
@@ -317,11 +324,45 @@ def move_pending_dispatches(database_path: Path) -> int:
     return len(conversations)
 
 
+#: Why a pre-cutover write job that may have reached HY was closed.
+ABANDONED_AT_CUTOVER = "abandoned_at_cutover"
+
+
+def settle_legacy_unknown_write_jobs(database_path: Path) -> int:
+    """Close write jobs the legacy pipeline left possibly-sent. Safe to repeat.
+
+    They ran under the legacy pipeline, so the SDK's `get_write_outcome` knows
+    none of them and the unknown reconcile can never settle them: a migrated
+    1.0.4 home showed 1000+ unknown forever. They are not retried, since they
+    may already have written and the SDK has no idempotency key. `cancelled`,
+    not `failed`: the upgrade ended them, nothing broke, and a `failed` count
+    would put the home in the failed state for good. `hy_invoked` is kept, so
+    no readmission path can pick one up again.
+    """
+    with main_database(database_path).read(row_factory=sqlite3.Row) as database:
+        if "memory_write_jobs" not in _tables(database):
+            return 0
+    now = datetime.now(UTC).isoformat()
+    with main_database(database_path).write(autocommit=True) as database:
+        # A `running` row already marked invoked would be reclaimed to
+        # `unknown` on the next open, so it is the same case.
+        return int(
+            database.execute(
+                "UPDATE memory_write_jobs SET state=?,lease_owner=NULL,"
+                "lease_expires_at=NULL,last_error=?,updated_at=? "
+                "WHERE state=? OR (state=? AND hy_invoked=1)",
+                (CANCELLED, ABANDONED_AT_CUTOVER, now, UNKNOWN, RUNNING),
+            ).rowcount
+        )
+
+
 def abandoned_legacy_work(database_path: Path) -> dict[str, int]:
     """Count legacy work that ends with the cutover instead of blocking it.
 
     A task that may have reached the SDK is not replayed: the SDK already holds
     whatever it committed. Pending user deletions are not carried over.
+    Settled write jobs are counted by their code, so a repeated preparation
+    reports the same number.
     """
     abandoned: dict[str, int] = {}
     with main_database(database_path).read(row_factory=sqlite3.Row) as database:
@@ -332,6 +373,11 @@ def abandoned_legacy_work(database_path: Path) -> dict[str, int]:
                 "memory_tasks",
                 "status='unknown' OR (status='running' AND hy_invocation_count > 0)",
             ),
+            (
+                "unknown_write_jobs",
+                "memory_write_jobs",
+                f"state='{CANCELLED}' AND last_error='{ABANDONED_AT_CUTOVER}'",
+            ),
             ("deletions", "memory_deletions", "status IN ('pending','running')"),
         )
         for name, table, where in checks:
@@ -354,6 +400,10 @@ def prepare_legacy_work_for_cutover(
     running, so a failure names it.
     """
     trace = {} if trace is None else trace
+    # First: a possibly-sent job never runs again, so its input is not
+    # frozen, and input released since must not block the cutover.
+    trace["sub_step"] = "unknown_write_jobs"
+    settle_legacy_unknown_write_jobs(database_path)
     trace["sub_step"] = "write_queue"
     frozen, moved = prepare_legacy_write_queue_for_cutover(database_path)
     trace["sub_step"] = "pending_dispatches"
@@ -411,13 +461,17 @@ def resume_legacy_cleanup(*, data_dir: Path, database_path: Path) -> tuple[str,
             bundle_id=marker.backup_bundle_id,
         )
     from one_memory.hy_runtime_config import HY_MEMORY_NAMESPACE
+    from one_memory.memory_visibility import drop_stale_triggers
 
     recover_abandoned_staging(
         data_dir=data_dir,
         database_path=database_path,
         namespace=HY_MEMORY_NAMESPACE,
     )
-    return drop_legacy_tables(database_path)
+    dropped = drop_legacy_tables(database_path)
+    # The visibility cache's grant triggers read a table just dropped (#233).
+    drop_stale_triggers(database_path)
+    return dropped
 
 
 MIGRATION_PREFIX = "sdk-authority-v2"
@@ -654,13 +708,14 @@ def run_startup_migration(
         step["authority"] = authority.value
     log.migration_id = authority_migration
     with log.step("preflight.profile_read") as step:
-        profile = read_active_profile(database_path)
-        step["profile_present"] = profile is not None
+        active = read_active_profile(database_path)
+        step["profile_present"] = active is not None
+    # No active profile means no generation model, not no embedding: the
+    # vectors are the built-in model's, which is also what the Runtime reads.
+    profile = active if active is not None else builtin_retrieval_snapshot()
     if authority is MemoryAuthority.SDK:
         with log.step("preflight.leftovers_scan") as step:
-            vectors_waiting = profile is not None and bool(
-                legacy_vector_collections(data_dir, profile)
-            )
+            vectors_waiting = bool(legacy_vector_collections(data_dir, profile))
             leftovers = bool(legacy_tables_present(database_path)) or abandoned_staging_present(
                 data_dir=data_dir, namespace=namespace
             )
@@ -671,7 +726,6 @@ def run_startup_migration(
         reattached: tuple[ReattachedCollection, ...] = ()
         readback: ReadbackReport | None = None
         if vectors_waiting:
-            assert profile is not None
             progress("vectors")
             with log.step("vectors.reattach") as step:
                 reattached = reattach_legacy_vectors(data_dir, profile)
@@ -723,12 +777,12 @@ def run_startup_migration(
         step["chat"] = len(legacy.chat)
         step["productivity"] = len(legacy.productivity)
         step["users"] = len(legacy.users)
-    if profile is None and not legacy.empty:
-        # Without the embedding profile the Runtime's collections cannot be
-        # named, so neither reattached nor read back. Legacy stays intact.
-        log.check("preflight.model_profile_check", "sdk_migration_model_profile_required")
-        raise DataPlaneMigrationError("sdk_migration_model_profile_required")
-    log.check("preflight.model_profile_check", None)
+    if active is None and not legacy.empty and vectors_belong_to_another_space(data_dir, profile):
+        # The memories live in the space of a profile that is no longer active,
+        # so the built-in model cannot read them. Legacy stays intact.
+        log.check("preflight.vector_space_check", "sdk_migration_vector_profile_missing")
+        raise DataPlaneMigrationError("sdk_migration_vector_profile_missing")
+    log.check("preflight.vector_space_check", None)
     with capture_fence(data_dir):
         # Builds through 19 left a staged SDK copy beside the original store.
         # Settling it before the bundle means the bundle holds the original
@@ -784,17 +838,23 @@ def run_startup_migration(
                 step.update(preparation.as_document())
             reattached = ()
             readback = None
-            if profile is not None:
-                progress("vectors")
-                with log.step("vectors.reattach") as step:
-                    reattached = reattach_legacy_vectors(data_dir, profile)
-                    step["collections"] = len(reattached)
-            if not legacy.empty:
-                assert profile is not None
+            # Before reattach touches the store, and after the staging
+            # recovery above has settled which store is the original.
+            with log.step("verify.stored_ids") as step:
+                expected, missing = expected_memory_ids(legacy, stored_memory_ids(data_dir))
+                step["chat_expected"] = len(expected.chat)
+                step["chat_missing_before_migration"] = missing
+            progress("vectors")
+            with log.step("vectors.reattach") as step:
+                reattached = reattach_legacy_vectors(data_dir, profile)
+                step["collections"] = len(reattached)
+            if not expected.empty:
                 progress("verify")
                 with log.step("verify.readback") as step:
                     with client_for(profile) as client:
-                        readback = verify_readback(client, legacy)
+                        readback = verify_readback(
+                            client, expected, chat_missing_before_migration=missing
+                        )
                     step.update(readback.as_document())
                     if not readback.passed:
                         store.set_phase(migration_id, phase="readback_failed", state="blocked")

memhub/apps/core/src/one_memory/sdk_migration_verify.py

@@ -14,12 +14,12 @@ import json
 import sqlite3
 from collections.abc import Iterator
 from contextlib import closing, contextmanager
-from dataclasses import dataclass
+from dataclasses import dataclass, replace
 from pathlib import Path
 from typing import Any
 
 from one_memory.database_access import main_database
-from one_memory.model_profiles import ModelProfileSnapshot
+from one_memory.model_profiles import ModelProfileSnapshot, pinned_embedding_base_url
 from one_memory.sdk_data_plane_migration import DataPlaneMigrationError
 
 _PRODUCTIVITY_KINDS = frozenset({"coding", "working"})
@@ -47,6 +47,8 @@ class ReadbackReport:
     chat_readable: int
     productivity_expected: int
     productivity_readable: int
+    #: Confirmed in the mirror, gone from the SDK before this migration began.
+    chat_missing_before_migration: int = 0
 
     @property
     def passed(self) -> bool:
@@ -63,6 +65,7 @@ class ReadbackReport:
             "chat_readable": self.chat_readable,
             "productivity_expected": self.productivity_expected,
             "productivity_readable": self.productivity_readable,
+            "chat_missing_before_migration": self.chat_missing_before_migration,
             "passed": self.passed,
         }
 
@@ -115,6 +118,27 @@ def read_legacy_memory_ids(connection: sqlite3.Connection) -> LegacyMemoryIds:
     )
 
 
+def expected_memory_ids(
+    legacy: LegacyMemoryIds, stored: frozenset[str]
+) -> tuple[LegacyMemoryIds, int]:
+    """What the readback must find, and how many confirmed chats it need not.
+
+    A 1.0.4 library can hold confirmed rows whose SDK node the SDK had already
+    merged or deleted: no migration could read those back (658 confirmed, 657
+    in the store, and the whole upgrade failed). So chat is held to what the
+    store held before this run, and anything in there must still read back.
+
+    If none of the confirmed chats is in the store, that is not history: the
+    store is empty or is the wrong one. Nothing is excused then, so the
+    readback still fails on every id instead of checking nothing just before
+    the mirror is dropped.
+    """
+    chat = tuple(memory_id for memory_id in legacy.chat if memory_id in stored)
+    if not chat:
+        return legacy, 0
+    return replace(legacy, chat=chat), len(legacy.chat) - len(chat)
+
+
 def legacy_memory_ids_from(database_path: Path) -> LegacyMemoryIds:
     with main_database(database_path).read() as connection:
         return read_legacy_memory_ids(connection)
@@ -160,7 +184,14 @@ def read_active_profile(database_path: Path) -> ModelProfileSnapshot | None:
         model=row["model"],
         embedding_model=row["embedding_model"],
         embedding_provider_kind=row["embedding_provider_kind"],
-        embedding_base_url=row["embedding_base_url"],
+        # Read-only, beside a Core that may not have pinned it yet.
+        embedding_base_url=pinned_embedding_base_url(
+            provider_kind=row["provider_kind"],
+            base_url=row["base_url"],
+            embedding_model=row["embedding_model"],
+            embedding_provider_kind=row["embedding_provider_kind"],
+            embedding_base_url=row["embedding_base_url"],
+        ),
         embedding_dimensions=row["embedding_dimensions"],
         timeout_seconds=row["timeout_seconds"],
         max_concurrency=row["max_concurrency"],
@@ -179,14 +210,18 @@ def runtime_sdk_client(
     from one_memory.model_profiles import ModelProfileStore
     from one_memory.secrets import KeyringSecretStore, SecretStoreUnavailableError
 
-    try:
-        secret = ModelProfileStore(database_path).resolve_snapshot_secret(
-            profile.profile_id, profile.revision, KeyringSecretStore()
-        )
-    except SecretStoreUnavailableError:
-        raise DataPlaneMigrationError("sdk_migration_secure_storage_unavailable") from None
-    if profile.has_api_key and secret is None:
-        raise DataPlaneMigrationError("sdk_migration_model_secret_unavailable")
+    secret: str | None = None
+    # Without a key there is nothing to resolve, and the built-in retrieval
+    # snapshot has no stored version to look one up by.
+    if profile.has_api_key:
+        try:
+            secret = ModelProfileStore(database_path).resolve_snapshot_secret(
+                profile.profile_id, profile.revision, KeyringSecretStore()
+            )
+        except SecretStoreUnavailableError:
+            raise DataPlaneMigrationError("sdk_migration_secure_storage_unavailable") from None
+        if secret is None:
+            raise DataPlaneMigrationError("sdk_migration_model_secret_unavailable")
     lease = acquire_client(profile, api_key=secret, data_dir=data_dir)
     try:
         yield lease.client
@@ -194,7 +229,9 @@ def runtime_sdk_client(
         lease.close()
 
 
-def verify_readback(client: Any, legacy: LegacyMemoryIds) -> ReadbackReport:
+def verify_readback(
+    client: Any, legacy: LegacyMemoryIds, *, chat_missing_before_migration: int = 0
+) -> ReadbackReport:
     listed: set[str] = set()
     for user_id in legacy.users or ("local",):
         offset = 0
@@ -228,5 +265,6 @@ def verify_readback(client: Any, legacy: LegacyMemoryIds) -> ReadbackReport:
         chat_readable=sum(1 for memory_id in legacy.chat if readable(memory_id)),
         productivity_expected=len(legacy.productivity),
         productivity_readable=sum(1 for memory_id in legacy.productivity if readable(memory_id)),
+        chat_missing_before_migration=chat_missing_before_migration,
     )
 

memhub/apps/core/src/one_memory/sdk_read_handle.py

@@ -0,0 +1,135 @@
+"""The read client, kept in the embedding space the active profile writes to.
+
+Every read used to open the built-in 512-dim collection while writes followed
+the active profile, so a custom embedding read an empty space (#242). This
+handle opens the active profile's `retrieval_view` instead, and the built-in
+space only when no profile is active.
+
+A profile change builds a new client and retires the old one; a retired client
+is closed once no task still uses it, as `SdkWriteCoordinator` does for writes
+(#218). A use is the task that asked for the client: callers keep their
+`client = await handle.client()` shape, and the use ends with that task (one
+HTTP request, one retention pass).
+
+The memory mode (lite / pro / ultra) is fixed per SDK client, so a mode change
+rebuilds and retires the same way, and so do the SDK features
+(`memory_features`): handles are keyed by (profile, mode, features). Neither
+mode nor features names a collection; both are separate client parameters.
+"""
+
+from __future__ import annotations
+
+import asyncio
+from collections.abc import Callable
+from pathlib import Path
+from typing import Any
+
+from one_memory.hy_graph_admin import effective_memory_mode
+from one_memory.hy_sdk_runtime import (
+    BUILTIN_RETRIEVAL_PROFILE_ID,
+    SdkClientHandle,
+    builtin_retrieval_snapshot,
+    retrieval_view,
+)
+from one_memory.hy_sdk_types import HySdkRuntimeError
+from one_memory.memory_features import FeatureOverrides
+from one_memory.model_profiles import ModelProfileStore, ProviderKind
+from one_memory.secrets import SecretStore, SecretStoreUnavailableError
+
+_BUILTIN = (BUILTIN_RETRIEVAL_PROFILE_ID, 0)
+
+
+class ActiveProfileReadHandle(SdkClientHandle):
+    """An `SdkClientHandle` whose profile follows the active one."""
+
+    def __init__(
+        self,
+        *,
+        profiles: ModelProfileStore,
+        secrets: SecretStore,
+        data_dir: Path,
+        memory_mode: Callable[[], str] = effective_memory_mode,
+        features: Callable[[], FeatureOverrides] = lambda: (),
+    ) -> None:
+        super().__init__(builtin_retrieval_snapshot(), api_key=None, data_dir=data_dir)
+        self.profiles = profiles
+        self.secrets = secrets
+        self.data_dir = data_dir
+        self.memory_mode = memory_mode
+        self.current_features = features
+        self._handles: dict[tuple[str, int, str, FeatureOverrides], SdkClientHandle] = {}
+        #: Tasks in flight per handle; a retired handle closes at zero.
+        self._uses: dict[SdkClientHandle, int] = {}
+        self._retired: set[SdkClientHandle] = set()
+        self._closing: set[asyncio.Task[None]] = set()
+        self._handle_lock = asyncio.Lock()
+
+    async def client(self) -> Any:
+        profile = await asyncio.to_thread(self.profiles.active_identity) or _BUILTIN
+        identity = (*profile, self.memory_mode(), self.current_features())
+        async with self._handle_lock:
+            handle = self._handles.get(identity)
+            if handle is None:
+                # Only on a change of profile: resolving a key can reach the
+                # system keychain, so it is not done on every read.
+                handle = await asyncio.to_thread(self._build, identity)
+                self._handles[identity] = handle
+            self._retired.update(value for key, value in self._handles.items() if key != identity)
+            self._handles = {identity: handle}
+            unused = {value for value in self._retired if not self._uses.get(value)}
+            self._retired -= unused
+            self._hold(handle)
+        await asyncio.gather(*(value.close() for value in unused))
+        return await handle.client()
+
+    async def close(self) -> None:
+        async with self._handle_lock:
+            handles = (*self._handles.values(), *self._retired)
+            self._handles, self._retired, self._uses = {}, set(), {}
+        await asyncio.gather(*(handle.close() for handle in handles), *self._closing)
+
+    def _hold(self, handle: SdkClientHandle) -> None:
+        """Count a use of `handle` until the asking task finishes."""
+        task = asyncio.current_task()
+        if task is None:  # pragma: no cover - always called from a coroutine
+            return
+        self._uses[handle] = self._uses.get(handle, 0) + 1
+        task.add_done_callback(lambda _task: self._release(handle))
+
+    def _release(self, handle: SdkClientHandle) -> None:
+        remaining = self._uses.get(handle, 0) - 1
+        if remaining > 0:
+            self._uses[handle] = remaining
+            return
+        self._uses.pop(handle, None)
+        if handle in self._retired:
+            self._retired.discard(handle)
+            closing = asyncio.get_running_loop().create_task(handle.close())
+            self._closing.add(closing)
+            closing.add_done_callback(self._closing.discard)
+
+    def _build(self, key: tuple[str, int, str, FeatureOverrides]) -> SdkClientHandle:
+        profile_id, revision, mode, features = key
+        identity = (profile_id, revision)
+        if identity == _BUILTIN:
+            return SdkClientHandle(
+                builtin_retrieval_snapshot(),
+                api_key=None,
+                data_dir=self.data_dir,
+                mode=mode,
+                features=features,
+            )
+        profile = self.profiles.snapshot(*identity)
+        view = retrieval_view(profile)
+        secret: str | None = None
+        # Only a remote OpenAI-compatible embedding sends the shared key.
+        if view.embedding_provider_kind is ProviderKind.OPENAI_COMPATIBLE:
+            try:
+                secret = self.profiles.resolve_snapshot_secret(*identity, self.secrets)
+            except SecretStoreUnavailableError:
+                raise HySdkRuntimeError("secure_storage_unavailable") from None
+            if secret is None:
+                raise HySdkRuntimeError("model_secret_unavailable")
+        return SdkClientHandle(
+            view, api_key=secret, data_dir=self.data_dir, mode=mode, features=features
+        )