diff --git a/CHANGELOG.md b/CHANGELOG.md
index 2ccecb6..af92329 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -4,6 +4,26 @@ Hy Memory and MemHub are versioned and released independently.
 
 ## Unreleased
 
+- 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 +88,44 @@ 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 2.0.8 — 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
diff --git a/docs/env_reference.md b/docs/env_reference.md
index f26caf9..7bd3841 100644
--- a/docs/env_reference.md
+++ b/docs/env_reference.md
@@ -191,6 +191,7 @@ 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 |
diff --git a/hy_memory/agent/extractor.py b/hy_memory/agent/extractor.py
index 8861600..5a81f44 100644
--- a/hy_memory/agent/extractor.py
+++ b/hy_memory/agent/extractor.py
@@ -1177,7 +1177,7 @@ 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 +1311,12 @@ 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,
diff --git a/hy_memory/agent/llm_provider.py b/hy_memory/agent/llm_provider.py
index 52a70b8..289f1ea 100644
--- a/hy_memory/agent/llm_provider.py
+++ b/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
diff --git a/hy_memory/agent/mem_agent.py b/hy_memory/agent/mem_agent.py
index 7c8db7b..e9ecb60 100644
--- a/hy_memory/agent/mem_agent.py
+++ b/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:
diff --git a/hy_memory/agent/reconciler.py b/hy_memory/agent/reconciler.py
index ce676ba..6c67cf0 100644
--- a/hy_memory/agent/reconciler.py
+++ b/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
diff --git a/hy_memory/agent/tools/base.py b/hy_memory/agent/tools/base.py
index 667bfc6..bf457b3 100644
--- a/hy_memory/agent/tools/base.py
+++ b/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
diff --git a/hy_memory/client.py b/hy_memory/client.py
index dd37ed3..8473917 100644
--- a/hy_memory/client.py
+++ b/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
@@ -4028,14 +4034,14 @@ 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 +4076,18 @@ 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 +4166,7 @@ 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 +4219,7 @@ 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 +4247,11 @@ 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 +5932,7 @@ 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,
diff --git a/hy_memory/coding/extractor.py b/hy_memory/coding/extractor.py
index 1d65cff..82afde6 100644
--- a/hy_memory/coding/extractor.py
+++ b/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
diff --git a/hy_memory/coding/fact_reconciler.py b/hy_memory/coding/fact_reconciler.py
index 91da943..9cc946e 100644
--- a/hy_memory/coding/fact_reconciler.py
+++ b/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:
diff --git a/hy_memory/coding/judge.py b/hy_memory/coding/judge.py
index bfa28c1..d005174 100644
--- a/hy_memory/coding/judge.py
+++ b/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
 
 
diff --git a/hy_memory/coding/reconciler.py b/hy_memory/coding/reconciler.py
index ed36dd6..7cfff82 100644
--- a/hy_memory/coding/reconciler.py
+++ b/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 ""
diff --git a/hy_memory/coding/store.py b/hy_memory/coding/store.py
index 5503934..0effea1 100644
--- a/hy_memory/coding/store.py
+++ b/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(),
                 ),
             )
diff --git a/hy_memory/coding/types.py b/hy_memory/coding/types.py
index 295be7a..147ddb6 100644
--- a/hy_memory/coding/types.py
+++ b/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,
         }
 
diff --git a/hy_memory/coding/writer.py b/hy_memory/coding/writer.py
index c6a8b1e..f4afcd3 100644
--- a/hy_memory/coding/writer.py
+++ b/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
diff --git a/hy_memory/core/embed_service.py b/hy_memory/core/embed_service.py
index efe7955..96fec8c 100644
--- a/hy_memory/core/embed_service.py
+++ b/hy_memory/core/embed_service.py
@@ -80,7 +80,7 @@ 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 +98,8 @@ 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
@@ -753,16 +754,22 @@ 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:
@@ -770,6 +777,10 @@ 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]
diff --git a/hy_memory/pipelines/_retrieval/reconcile_retrieval.py b/hy_memory/pipelines/_retrieval/reconcile_retrieval.py
index a5081e1..526aa0d 100644
--- a/hy_memory/pipelines/_retrieval/reconcile_retrieval.py
+++ b/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}"
         )
 
diff --git a/hy_memory/pipelines/cross_domain_sweeper.py b/hy_memory/pipelines/cross_domain_sweeper.py
index 700423e..8843e40 100644
--- a/hy_memory/pipelines/cross_domain_sweeper.py
+++ b/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:
diff --git a/hy_memory/pipelines/reader_base.py b/hy_memory/pipelines/reader_base.py
index bc288d4..8bcacef 100644
--- a/hy_memory/pipelines/reader_base.py
+++ b/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
 
diff --git a/hy_memory/pipelines/system2_agent.py b/hy_memory/pipelines/system2_agent.py
index d16d242..2ace7a8 100644
--- a/hy_memory/pipelines/system2_agent.py
+++ b/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
 
 
diff --git a/hy_memory/utils/async_executor.py b/hy_memory/utils/async_executor.py
index 0ab0c7e..28afa74 100644
--- a/hy_memory/utils/async_executor.py
+++ b/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(
diff --git a/hy_memory/utils/pipeline_log_writer.py b/hy_memory/utils/pipeline_log_writer.py
index 20220bb..7b3f1b7 100644
--- a/hy_memory/utils/pipeline_log_writer.py
+++ b/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,
diff --git a/hy_memory/utils/pipeline_observability.py b/hy_memory/utils/pipeline_observability.py
index a7d898e..371dc70 100644
--- a/hy_memory/utils/pipeline_observability.py
+++ b/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,
diff --git a/hy_memory/utils/token_count.py b/hy_memory/utils/token_count.py
index 10e53ad..afbf09d 100644
--- a/hy_memory/utils/token_count.py
+++ b/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)
diff --git a/memhub/PRIVACY.md b/memhub/PRIVACY.md
index bae6628..e4616df 100644
--- a/memhub/PRIVACY.md
+++ b/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
diff --git a/memhub/README.md b/memhub/README.md
index 0c06ec2..6f609e6 100644
--- a/memhub/README.md
+++ b/memhub/README.md
@@ -18,7 +18,7 @@ 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`
+源码当前固定使用公开 PyPI 正式版 `hy-memory==2.0.8`。MemHub 源码版本为 `1.0.5`
 （build 14），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
@@ -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==2.0.8`。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`。
 
diff --git a/memhub/SBOM.spdx.json b/memhub/SBOM.spdx.json
index f5f91d3..cff821a 100644
--- a/memhub/SBOM.spdx.json
+++ b/memhub/SBOM.spdx.json
@@ -2352,7 +2352,7 @@
     {
       "SPDXID": "SPDXRef-Package-hy-memory",
       "name": "hy-memory",
-      "versionInfo": "2.0.6",
+      "versionInfo": "2.0.8",
       "downloadLocation": "NOASSERTION",
       "filesAnalyzed": false,
       "licenseConcluded": "Apache-2.0",
diff --git a/memhub/THIRD_PARTY_NOTICES.md b/memhub/THIRD_PARTY_NOTICES.md
index 58c3ab6..58b9923 100644
--- a/memhub/THIRD_PARTY_NOTICES.md
+++ b/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 2.0.8
 
 License: Apache-2.0
 
diff --git a/memhub/apps/core/pyproject.toml b/memhub/apps/core/pyproject.toml
index ea2dce7..c36f94d 100644
--- a/memhub/apps/core/pyproject.toml
+++ b/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==2.0.8",
     # 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",
diff --git a/memhub/apps/core/src/one_memory/api.py b/memhub/apps/core/src/one_memory/api.py
index 255c353..a674d7e 100644
--- a/memhub/apps/core/src/one_memory/api.py
+++ b/memhub/apps/core/src/one_memory/api.py
@@ -3175,25 +3175,25 @@ def create_app(
             raise HTTPException(status_code=422, detail="memory_task_requires_experience")
         namespace = _namespace_kwargs(request.kind, request.variant)
         try:
-            client = await app.state.sdk_write_coordinator.client()
-            update_kwargs: dict[str, object] = dict(namespace)
-            if request.task is not None:
-                update_kwargs["task"] = request.task
-            result = await asyncio.to_thread(
-                client.update, memory_id, request.content, **update_kwargs
-            )
-            if not isinstance(result, dict):
-                raise HySdkError("hy_memory_update_rejected")
-            if result.get("success") is not True:
-                if "not found" in str(result.get("error_message") or ""):
-                    raise HTTPException(status_code=404, detail="memory_not_found")
-                raise HySdkError("hy_memory_update_rejected")
-            # The SDK names the namespace it wrote; read that one back.
-            if result.get("kind") in {"experience", "fact"}:
-                namespace = {"kind": result["kind"]}
-                if result.get("variant") in {"working", "coding"}:
-                    namespace["variant"] = result["variant"]
-            raw = await asyncio.to_thread(client.get, memory_id, **namespace)
+            async with app.state.sdk_write_coordinator.use_client() as client:
+                update_kwargs: dict[str, object] = dict(namespace)
+                if request.task is not None:
+                    update_kwargs["task"] = request.task
+                result = await asyncio.to_thread(
+                    client.update, memory_id, request.content, **update_kwargs
+                )
+                if not isinstance(result, dict):
+                    raise HySdkError("hy_memory_update_rejected")
+                if result.get("success") is not True:
+                    if "not found" in str(result.get("error_message") or ""):
+                        raise HTTPException(status_code=404, detail="memory_not_found")
+                    raise HySdkError("hy_memory_update_rejected")
+                # The SDK names the namespace it wrote; read that one back.
+                if result.get("kind") in {"experience", "fact"}:
+                    namespace = {"kind": result["kind"]}
+                    if result.get("variant") in {"working", "coding"}:
+                        namespace["variant"] = result["variant"]
+                raw = await asyncio.to_thread(client.get, memory_id, **namespace)
             memory = None if raw is None else HySdkMemoryRecord.from_sdk(raw)
         except HTTPException:
             raise
@@ -3222,10 +3222,10 @@ def create_app(
     ) -> Response:
         _check_memory_namespace(kind, variant)
         try:
-            client = await app.state.sdk_write_coordinator.client()
-            result = await asyncio.to_thread(
-                client.delete, memory_id, **_namespace_kwargs(kind, variant)
-            )
+            async with app.state.sdk_write_coordinator.use_client() as client:
+                result = await asyncio.to_thread(
+                    client.delete, memory_id, **_namespace_kwargs(kind, variant)
+                )
             if not isinstance(result, dict):
                 raise HySdkError("hy_memory_delete_rejected")
             error_code = result.get("error_code")
@@ -3467,9 +3467,9 @@ def create_app(
         """Retry unknown writes only behind the user's explicit confirmation."""
         del request
         reconciled = await _sdk_memory_runtime(app).reconcile_unknown_outputs()
-        from one_memory.memory_server.durable_queue import DurableWriteQueue
-
-        queue = DurableWriteQueue(app.state.memory_write_queue_path)
+        # The startup queue knows the authority. A fresh legacy-mode queue
+        # would recreate the job-evidence table the SDK cutover dropped.
+        queue = app.state.memory_write_queue
         unknown = queue.unknown_jobs(limit=2048)
         if app.state.memory_authority is MemoryAuthority.LEGACY:
             for job in unknown:
diff --git a/memhub/apps/core/src/one_memory/hy_runtime_config.py b/memhub/apps/core/src/one_memory/hy_runtime_config.py
index 097e450..cef9958 100644
--- a/memhub/apps/core/src/one_memory/hy_runtime_config.py
+++ b/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 = "2.0.8"
 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"
diff --git a/memhub/apps/core/src/one_memory/hy_sdk_migration.py b/memhub/apps/core/src/one_memory/hy_sdk_migration.py
index cf43a29..e111eac 100644
--- a/memhub/apps/core/src/one_memory/hy_sdk_migration.py
+++ b/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==2.0.8 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])
diff --git a/memhub/apps/core/src/one_memory/hy_sdk_runtime.py b/memhub/apps/core/src/one_memory/hy_sdk_runtime.py
index 854bb48..3b4f626 100644
--- a/memhub/apps/core/src/one_memory/hy_sdk_runtime.py
+++ b/memhub/apps/core/src/one_memory/hy_sdk_runtime.py
@@ -36,6 +36,7 @@ from one_memory.hy_graph_admin import (
 )
 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.model_profiles import (
     ModelProfileSnapshot,
     ProviderKind,
@@ -66,6 +67,16 @@ _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."""
@@ -381,16 +392,23 @@ 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="ultra" if graph_store_enabled() else "pro",
+                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)
 
 
diff --git a/memhub/apps/core/src/one_memory/hy_sdk_types.py b/memhub/apps/core/src/one_memory/hy_sdk_types.py
index 52fd381..d510a79 100644
--- a/memhub/apps/core/src/one_memory/hy_sdk_types.py
+++ b/memhub/apps/core/src/one_memory/hy_sdk_types.py
@@ -151,6 +151,7 @@ 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 +168,9 @@ 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),
diff --git a/memhub/apps/core/src/one_memory/log_file.py b/memhub/apps/core/src/one_memory/log_file.py
index 23be281..27b207b 100644
--- a/memhub/apps/core/src/one_memory/log_file.py
+++ b/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))
diff --git a/memhub/apps/core/src/one_memory/memory_server/durable_queue.py b/memhub/apps/core/src/one_memory/memory_server/durable_queue.py
index b755ba9..5d74014 100644
--- a/memhub/apps/core/src/one_memory/memory_server/durable_queue.py
+++ b/memhub/apps/core/src/one_memory/memory_server/durable_queue.py
@@ -52,6 +52,7 @@ 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
 
 logger = logging.getLogger(__name__)
@@ -241,14 +242,16 @@ 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 +295,23 @@ _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
@@ -1982,21 +2002,24 @@ 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 +2121,7 @@ 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 +2179,7 @@ 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 +2196,14 @@ 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
@@ -2559,9 +2586,22 @@ 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,
diff --git a/memhub/apps/core/src/one_memory/sdk_memory_runtime.py b/memhub/apps/core/src/one_memory/sdk_memory_runtime.py
index fd247e5..4e13689 100644
--- a/memhub/apps/core/src/one_memory/sdk_memory_runtime.py
+++ b/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
+from contextlib import AsyncExitStack, asynccontextmanager, suppress
 from datetime import UTC, datetime, timedelta
 from pathlib import Path
 from typing import Any
@@ -164,44 +165,45 @@ class SdkMemoryRuntime:
             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,
+                "local",
+                "*",
+                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),
@@ -240,36 +242,39 @@ class SdkMemoryRuntime:
             decision=decision,
         )
 
-    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")
diff --git a/memhub/apps/core/src/one_memory/sdk_migration_runner.py b/memhub/apps/core/src/one_memory/sdk_migration_runner.py
index 203eddc..a98ef46 100644
--- a/memhub/apps/core/src/one_memory/sdk_migration_runner.py
+++ b/memhub/apps/core/src/one_memory/sdk_migration_runner.py
@@ -35,6 +35,8 @@ 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
@@ -317,11 +319,45 @@ 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 +368,11 @@ 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 +395,10 @@ 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"
diff --git a/memhub/apps/core/src/one_memory/sdk_write_processor.py b/memhub/apps/core/src/one_memory/sdk_write_processor.py
index f6e4d98..dd5d94a 100644
--- a/memhub/apps/core/src/one_memory/sdk_write_processor.py
+++ b/memhub/apps/core/src/one_memory/sdk_write_processor.py
@@ -3,7 +3,8 @@
 from __future__ import annotations
 
 import asyncio
-from collections.abc import Awaitable, Callable, Mapping, Sequence
+from collections.abc import AsyncIterator, Awaitable, Callable, Mapping, Sequence
+from contextlib import asynccontextmanager
 from dataclasses import dataclass
 from datetime import datetime
 from typing import Any
@@ -56,7 +57,12 @@ class SdkWriteProcessor:
 
         scope = _write_scope(records)
         metadata = _write_metadata(records, scope)
-        client = await self._client_handle.client()
+        try:
+            client = await self._client_handle.client()
+        except Exception as error:
+            # Before `before_hy_invocation`: nothing was sent, so the job
+            # waits and retries instead of being left `unknown`.
+            raise RuntimeError("hy_memory_client_unavailable") from error
 
         if extracted:
             contents = [sanitize_model_input(record.verbatim) for record in records]
@@ -125,7 +131,14 @@ class SdkWriteProcessor:
 
 
 class SdkWriteCoordinator:
-    """Resolve the active model profile, then execute one public SDK write."""
+    """Resolve the active model profile, then execute one public SDK write.
+
+    A model profile change retires the previous profile's handle, but only
+    closes it once nothing uses it. Closing it at once tore the client out
+    from under a write still running on it: the SDK cleared its stores and
+    the SharedRuntime closed mid-`add`, the job failed with AttributeError
+    after HY was invoked, and it was left `unknown`.
+    """
 
     def __init__(
         self,
@@ -142,6 +155,9 @@ class SdkWriteCoordinator:
         self.transport_factory = transport_factory
         self.sedimentation_enabled = sedimentation_enabled
         self._handles: dict[tuple[str, int], SdkClientHandle] = {}
+        #: Uses in flight per handle; a retired handle closes at zero.
+        self._uses: dict[SdkClientHandle, int] = {}
+        self._retired: set[SdkClientHandle] = set()
         self._lock = asyncio.Lock()
 
     async def process_sdk_job(
@@ -153,28 +169,30 @@ class SdkWriteCoordinator:
     ) -> SdkWriteReceipt:
         if not self.sedimentation_enabled():
             raise RuntimeError("model_sedimentation_disabled")
-        handle = await self._active_handle()
-        return await SdkWriteProcessor(handle).process_sdk_job(
-            records,
-            request_id=request_id,
-            before_hy_invocation=before_hy_invocation,
-        )
+        async with self._active_handle() as handle:
+            return await SdkWriteProcessor(handle).process_sdk_job(
+                records,
+                request_id=request_id,
+                before_hy_invocation=before_hy_invocation,
+            )
 
     async def inquire_write_outcome(
         self, request_id: str
     ) -> SdkWriteReceipt | None:
-        return await SdkWriteProcessor(
-            await self._active_handle()
-        ).inquire_write_outcome(request_id)
+        async with self._active_handle() as handle:
+            return await SdkWriteProcessor(handle).inquire_write_outcome(request_id)
 
     async def close(self) -> None:
         async with self._lock:
-            handles, self._handles = tuple(self._handles.values()), {}
+            handles = (*self._handles.values(), *self._retired)
+            self._handles, self._retired, self._uses = {}, set(), {}
         await asyncio.gather(*(handle.close() for handle in handles))
 
-    async def client(self) -> Any:
-        """Return the public SDK client selected by the active write profile."""
-        return await (await self._active_handle()).client()
+    @asynccontextmanager
+    async def use_client(self) -> AsyncIterator[Any]:
+        """The public SDK client of the active write profile, kept open while used."""
+        async with self._active_handle() as handle:
+            yield await handle.client()
 
     async def active_profile_credentials(
         self,
@@ -183,7 +201,9 @@ class SdkWriteCoordinator:
         _identity, profile, secret = await asyncio.to_thread(self._read_active_credentials)
         return profile, secret
 
-    async def _active_handle(self) -> SdkClientHandle:
+    @asynccontextmanager
+    async def _active_handle(self) -> AsyncIterator[SdkClientHandle]:
+        """Use the active profile's handle; retire the others, closing the unused."""
         identity, profile, secret = await asyncio.to_thread(
             self._read_active_credentials
         )
@@ -197,14 +217,25 @@ class SdkWriteCoordinator:
                     http_transport_factory=self.transport_factory(profile),
                 )
                 self._handles[identity] = handle
-            stale = [
-                value
-                for key, value in self._handles.items()
-                if key != identity
-            ]
+            self._retired.update(value for key, value in self._handles.items() if key != identity)
             self._handles = {identity: handle}
-        await asyncio.gather(*(value.close() for value in stale))
-        return handle
+            self._uses[handle] = self._uses.get(handle, 0) + 1
+            unused = {value for value in self._retired if not self._uses.get(value)}
+            self._retired -= unused
+        await asyncio.gather(*(value.close() for value in unused))
+        try:
+            yield handle
+        finally:
+            # Bookkeeping without an await, so a cancelled caller still
+            # releases its use.
+            remaining = self._uses.get(handle, 0) - 1
+            if remaining > 0:
+                self._uses[handle] = remaining
+            else:
+                self._uses.pop(handle, None)
+            if remaining <= 0 and handle in self._retired:
+                self._retired.discard(handle)
+                await handle.close()
 
     def _read_active_credentials(
         self,
diff --git a/memhub/apps/core/tests/TEST_INDEX.md b/memhub/apps/core/tests/TEST_INDEX.md
index fd2f457..b38c5bb 100644
--- a/memhub/apps/core/tests/TEST_INDEX.md
+++ b/memhub/apps/core/tests/TEST_INDEX.md
@@ -79,7 +79,7 @@
 | `test_database_file_permissions.py` | `contract` | 覆盖 a new main database is readable by its owner alone、the wal sidecars are not left world readable、a database from an older build is tightened on open 场景。 |
 | `test_db_session_utc.py` | `contract` | The MySQL session must be pinned to UTC. |
 | `test_derived_echo.py` | `contract` | A host that saves our injection back must not mint a new user decision. |
-| `test_detailed_logging.py` | `contract` | 覆盖 detailed pipeline logging defaults off and survives restart、live log level applies to full hy and provider call chain、info collects hy info without enabling debug pipeline content 等 7 个场景。 |
+| `test_detailed_logging.py` | `contract` | 覆盖 detailed logging defaults off and survives restart、hy pipeline log and trace stay off at every level、live level change reaches the hy text log file 等 7 个场景。 |
 | `test_diagnostic_log.py` | `contract` | 覆盖 high frequency success events are summarized but warnings are not、structured log is private correlated and owner only、arbitrary detail and absolute stack are rejected 等 14 个场景。 |
 | `test_diagnostic_log_durability.py` | `contract` | 写线程报告的「已落盘」必须是真的。 |
 | `test_diagnostic_log_off_loop.py` | `contract` | 诊断日志不得在调用线程上做磁盘 I/O。 |
@@ -112,6 +112,7 @@
 | `test_hy_log_privacy.py` | `contract` | HY SDK records are not rewritten by One Memory's logging setup. |
 | `test_hy_network_egress.py` | `contract` | Where a real HY client is allowed to send bytes, and where it is not. |
 | `test_hy_runtime_config.py` | `contract` | MEMORY.md is the audited record of what we send HY; keep it answerable. |
+| `test_hy_sdk_log_routing.py` | `contract` | HY SDK text logs go to hy_memory.log only, with warnings also in core.log. |
 | `test_hy_step_trace_switch.py` | `contract` | Rule 15: the per-step HY trace is off unless the debug tier asks for it. |
 | `test_import_api.py` | `integration` | 覆盖 resume during cleanup returns conflict without internal error、authorize detect import and revoke api flow、protected api rejects missing token and foreign origin 等 5 个场景。 |
 | `test_import_control_races_review.py` | `contract` | Read-only review repro: real runner/store, synthetic delayed adapter cleanup. |
@@ -245,6 +246,7 @@
 | `test_schema_gate.py` | `contract` | 覆盖 fresh database records the build schema、future schema refuses to open、legacy database without a ledger is left untouched 场景。 |
 | `test_schema_migrations.py` | `contract` | 覆盖 evidence store adds source ordinal to previous schema、import store migrates previous job and error schema、old active job index lets a sync sit beside paused history 等 4 个场景。 |
 | `test_schema_ownership_gate.py` | `contract` | A1-03/A1-04: schema ownership, and the hard part of telling DDL from reads. |
+| `test_sdk_client_construction.py` | `contract` | SDK clients of one home are constructed one at a time. |
 | `test_sdk_data_plane_migration.py` | `contract` | SDK-authority migration ledger, aliases and atomic cutover marker. |
 | `test_sdk_memory_qa.py` | `contract` | 覆盖 sdk ask propagates sdk search failure without legacy fallback、sdk ask revalidates memory before returning extract、sdk ask uses native sdk memory without governance fields 场景。 |
 | `test_sdk_migration_bundle.py` | `contract` | SDK-authority cutover requires a verified management + SDK data-plane bundle. |
@@ -307,4 +309,4 @@
 | `test_workspace_recall.py` | `contract` | Workspace memories were unreachable from every host; these pin the way in. |
 | `test_write_queue_lease_renewal.py` | `contract` | Rule 8: a lease must outlive the work it covers, and be renewed while it runs. |
 
-共 302 个测试文件。
+共 304 个测试文件。
diff --git a/memhub/apps/core/tests/core_test_classification.py b/memhub/apps/core/tests/core_test_classification.py
index 5f362a3..7c0b208 100644
--- a/memhub/apps/core/tests/core_test_classification.py
+++ b/memhub/apps/core/tests/core_test_classification.py
@@ -14,8 +14,8 @@ from pathlib import Path
 import pytest
 
 PRIMARY_MARKERS = ("unit", "contract", "integration", "performance", "live")
-EXPECTED_TEST_FILE_COUNT = 302
-EXPECTED_TEST_FILES_SHA256 = "07707528025ac134e46c93ebef12335e852df1ed0f52f8beb62abb3b4d1b0e97"
+EXPECTED_TEST_FILE_COUNT = 304
+EXPECTED_TEST_FILES_SHA256 = "b28f022b9688c96ca563c59de63e95d344d2f01ba3b6020b422bf4dfb27520d6"
 
 # Pure transformations, validation models, parity checks and state machines.
 # This list is intentionally small: uncertain files default to contract.
diff --git a/memhub/apps/core/tests/test_detailed_logging.py b/memhub/apps/core/tests/test_detailed_logging.py
index 6bcdf44..dd61599 100644
--- a/memhub/apps/core/tests/test_detailed_logging.py
+++ b/memhub/apps/core/tests/test_detailed_logging.py
@@ -1,6 +1,5 @@
 import json
 import logging
-from importlib import import_module
 from pathlib import Path
 
 from one_memory.hy_runtime_config import (
@@ -12,20 +11,50 @@ from one_memory.hy_runtime_config import (
 )
 
 
-def test_detailed_pipeline_logging_defaults_off_and_survives_restart(
-    tmp_path: Path, monkeypatch
-) -> None:
-    monkeypatch.delenv("MEMORY_PIPELINE_TRACE_ENABLED", raising=False)
+def test_detailed_logging_defaults_off_and_survives_restart(tmp_path: Path, monkeypatch) -> None:
+    monkeypatch.delenv("MEMORY_LOG_LEVEL", raising=False)
     assert detailed_logging_enabled(tmp_path) is False
     configure_hy_runtime_environment(tmp_path)
-    assert __import__("os").environ["MEMORY_PIPELINE_TRACE_ENABLED"] == "false"
+    assert __import__("os").environ["MEMORY_LOG_LEVEL"] == "INFO"
     assert diagnostic_log_level(tmp_path) == "info"
 
     assert set_detailed_logging(tmp_path, True) is True
     assert detailed_logging_enabled(tmp_path) is True
-    monkeypatch.delenv("MEMORY_PIPELINE_TRACE_ENABLED", raising=False)
+    monkeypatch.delenv("MEMORY_LOG_LEVEL", raising=False)
     configure_hy_runtime_environment(tmp_path)
-    assert __import__("os").environ["MEMORY_PIPELINE_TRACE_ENABLED"] == "true"
+    assert __import__("os").environ["MEMORY_LOG_LEVEL"] == "DEBUG"
+
+
+def test_hy_pipeline_log_and_trace_stay_off_at_every_level(tmp_path: Path, monkeypatch) -> None:
+    environment = __import__("os").environ
+    for level in ("error", "warning", "info", "debug"):
+        for name in ("MEMORY_PIPELINE_LOG_ENABLED", "MEMORY_PIPELINE_TRACE_ENABLED"):
+            monkeypatch.setenv(name, "true")
+        set_diagnostic_log_level(tmp_path, level)
+        assert environment["MEMORY_PIPELINE_LOG_ENABLED"] == "false"
+        assert environment["MEMORY_PIPELINE_TRACE_ENABLED"] == "false"
+
+        for name in ("MEMORY_PIPELINE_LOG_ENABLED", "MEMORY_PIPELINE_TRACE_ENABLED"):
+            monkeypatch.setenv(name, "true")
+        configure_hy_runtime_environment(tmp_path)
+        assert environment["MEMORY_PIPELINE_LOG_ENABLED"] == "false"
+        assert environment["MEMORY_PIPELINE_TRACE_ENABLED"] == "false"
+
+
+def test_live_level_change_reaches_the_hy_text_log_file(tmp_path: Path) -> None:
+    sdk = logging.getLogger("hy_memory")
+    text_log = logging.FileHandler(tmp_path / "hy_memory.log", encoding="utf-8")
+    text_log.setLevel(logging.INFO)
+    sdk.addHandler(text_log)
+    try:
+        set_diagnostic_log_level(tmp_path, "debug")
+        assert text_log.level == logging.DEBUG
+        set_diagnostic_log_level(tmp_path, "warning")
+        assert text_log.level == logging.WARNING
+    finally:
+        sdk.removeHandler(text_log)
+        text_log.close()
+        set_diagnostic_log_level(tmp_path, "info")
 
 
 def test_live_log_level_applies_to_full_hy_and_provider_call_chain(
@@ -60,43 +89,9 @@ def test_info_collects_hy_info_without_enabling_debug_pipeline_content(
     assert logging.getLogger("openai._base_client").isEnabledFor(logging.INFO)
     assert not logging.getLogger("openai._base_client").isEnabledFor(logging.DEBUG)
     assert __import__("os").environ["MEMORY_LOG_LEVEL"] == "INFO"
-    assert __import__("os").environ["MEMORY_PIPELINE_TRACE_ENABLED"] == "false"
     assert __import__("os").environ["MEMORY_LOG_LLM_CONTENT"] == "false"
 
 
-def test_live_debug_switch_makes_sdk_pipeline_writer_keep_original_content(
-    tmp_path: Path,
-) -> None:
-    writer_type = import_module("hy_memory.utils.pipeline_log_writer").PipelineLogWriter
-    writer = writer_type(str(tmp_path / "pipeline"))
-
-    set_diagnostic_log_level(tmp_path, "info")
-    writer.write_step(
-        request_id="info",
-        step="EXTRACT",
-        prompt="info prompt",
-        response="info response",
-    )
-    set_diagnostic_log_level(tmp_path, "debug")
-    writer.write_step(
-        request_id="debug",
-        step="EXTRACT",
-        prompt="debug prompt",
-        response="debug response",
-    )
-
-    lines = [
-        json.loads(line)
-        for target in (tmp_path / "pipeline").rglob("*.log")
-        for line in target.read_text(encoding="utf-8").splitlines()
-    ]
-    info, debug = lines
-    assert info["prompt"].startswith("<redacted len=")
-    assert info["content"].startswith("<redacted len=")
-    assert debug["prompt"] == "debug prompt"
-    assert debug["content"] == "debug response"
-
-
 def test_log_level_document_migrates_the_old_boolean_shape(tmp_path: Path) -> None:
     target = tmp_path / "settings" / "diagnostic-logging.json"
     target.parent.mkdir(parents=True)
@@ -117,6 +112,7 @@ def test_configure_environment_passes_debug_content_settings_to_sdk(
         "MEMORY_DATA_DIR",
         "MEMORY_LOG_DIR",
         "MEMORY_LOG_LEVEL",
+        "MEMORY_PIPELINE_LOG_ENABLED",
         "MEMORY_PIPELINE_TRACE_ENABLED",
         "MEMORY_LOG_LLM_CONTENT",
     ):
@@ -129,5 +125,6 @@ def test_configure_environment_passes_debug_content_settings_to_sdk(
     assert environment["MEMORY_DATA_DIR"] == str(memory_root)
     assert environment["MEMORY_LOG_DIR"] == str(memory_root / "logs")
     assert environment["MEMORY_LOG_LEVEL"] == "DEBUG"
-    assert environment["MEMORY_PIPELINE_TRACE_ENABLED"] == "true"
+    assert environment["MEMORY_PIPELINE_LOG_ENABLED"] == "false"
+    assert environment["MEMORY_PIPELINE_TRACE_ENABLED"] == "false"
     assert environment["MEMORY_LOG_LLM_CONTENT"] == "true"
diff --git a/memhub/apps/core/tests/test_hy_sdk_log_routing.py b/memhub/apps/core/tests/test_hy_sdk_log_routing.py
new file mode 100644
index 0000000..4e8cb55
--- /dev/null
+++ b/memhub/apps/core/tests/test_hy_sdk_log_routing.py
@@ -0,0 +1,146 @@
+"""HY SDK text logs go to hy_memory.log only, with warnings also in core.log.
+
+The SDK's first client adds a stdout and a stderr stream to the "hy_memory"
+logger and stops it propagating. Under launchd those streams are
+service.log and service-error.log, so each SDK line used to be written
+twice and none reached core.log.
+"""
+
+import logging
+import sys
+from logging.handlers import RotatingFileHandler
+from pathlib import Path
+from types import SimpleNamespace
+
+import pytest
+
+from one_memory import hy_sdk_runtime
+from one_memory.log_file import (
+    HY_SDK_LOGGER,
+    confine_hy_sdk_logging,
+    log_path,
+    start_logging_to_file,
+)
+
+
+@pytest.fixture
+def sdk_logger():
+    """The "hy_memory" logger as HY's setup_logging leaves it, restored after."""
+    sdk = logging.getLogger(HY_SDK_LOGGER)
+    root = logging.getLogger()
+    saved = (list(sdk.handlers), sdk.propagate, sdk.level, list(root.handlers), root.level)
+    yield sdk
+    for handler in sdk.handlers:
+        if handler not in saved[0]:
+            handler.close()
+    for handler in root.handlers:
+        if handler not in saved[3]:
+            handler.close()
+    sdk.handlers, sdk.propagate, root.handlers = saved[0], saved[1], saved[3]
+    sdk.setLevel(saved[2])
+    root.setLevel(saved[4])
+
+
+def _configure_like_hy(sdk: logging.Logger, log_dir: Path) -> logging.FileHandler:
+    log_dir.mkdir(parents=True, exist_ok=True)
+    text_log = logging.FileHandler(log_dir / "hy_memory.log", encoding="utf-8")
+    stdout = logging.StreamHandler(sys.stdout)
+    stderr = logging.StreamHandler(sys.stderr)
+    stderr.setLevel(logging.WARNING)
+    sdk.handlers = [text_log, stdout, stderr]
+    sdk.propagate = False
+    sdk.setLevel(logging.INFO)
+    return text_log
+
+
+def test_sdk_lines_stay_in_hy_memory_log_and_warnings_reach_core_log(
+    tmp_path: Path, sdk_logger: logging.Logger
+) -> None:
+    start_logging_to_file(tmp_path)
+    text_log = _configure_like_hy(sdk_logger, tmp_path / "hy")
+
+    confine_hy_sdk_logging()
+    confine_hy_sdk_logging()
+
+    assert not [
+        handler
+        for handler in sdk_logger.handlers
+        if isinstance(handler, logging.StreamHandler)
+        and not isinstance(handler, logging.FileHandler)
+    ]
+    assert text_log in sdk_logger.handlers
+    assert len(sdk_logger.handlers) == 2, "the core.log forwarder is added once"
+
+    child = logging.getLogger("hy_memory.client")
+    child.info("[add] routine sdk line")
+    child.warning("[add] sdk problem")
+    for handler in logging.getLogger().handlers:
+        handler.flush()
+    text_log.flush()
+
+    hy_text = (tmp_path / "hy" / "hy_memory.log").read_text(encoding="utf-8")
+    core_text = log_path(tmp_path).read_text(encoding="utf-8")
+    assert "routine sdk line" in hy_text and "sdk problem" in hy_text
+    assert "sdk problem" in core_text
+    assert "routine sdk line" not in core_text
+    assert core_text.count("sdk problem") == 1
+
+
+def test_nothing_changes_before_hy_has_configured_its_logger(
+    tmp_path: Path, sdk_logger: logging.Logger
+) -> None:
+    start_logging_to_file(tmp_path)
+    sdk_logger.handlers = []
+    sdk_logger.propagate = True
+
+    confine_hy_sdk_logging()
+
+    assert sdk_logger.handlers == []
+
+
+def test_without_core_log_the_streams_are_still_dropped(
+    tmp_path: Path, sdk_logger: logging.Logger
+) -> None:
+    logging.getLogger().handlers = [
+        handler
+        for handler in logging.getLogger().handlers
+        if not isinstance(handler, RotatingFileHandler)
+    ]
+    text_log = _configure_like_hy(sdk_logger, tmp_path / "hy")
+
+    confine_hy_sdk_logging()
+
+    assert sdk_logger.handlers == [text_log]
+
+
+@pytest.mark.parametrize("fails", [False, True])
+def test_every_client_construction_confines_the_sdk_log(
+    tmp_path: Path, monkeypatch, fails: bool
+) -> None:
+    order: list[str] = []
+
+    class Client:
+        def __init__(self, **_options):
+            order.append("client")
+            if fails:
+                raise RuntimeError("construction failed")
+
+    monkeypatch.setattr(hy_sdk_runtime, "HyMemoryClient", Client)
+    monkeypatch.setattr(hy_sdk_runtime, "build_sdk_config", lambda *_a, **_k: SimpleNamespace())
+    monkeypatch.setattr(hy_sdk_runtime, "inspect_graph_width_mismatch", lambda _c: None)
+    monkeypatch.setattr(hy_sdk_runtime, "_acquire_runtime", lambda _c, _k: object())
+    monkeypatch.setattr(hy_sdk_runtime, "_release_runtime", lambda _k: None)
+    monkeypatch.setattr(hy_sdk_runtime, "confine_hy_sdk_logging", lambda: order.append("confine"))
+    profile = SimpleNamespace(embedding_provider_kind=None, provider_kind=None)
+
+    if fails:
+        with pytest.raises(RuntimeError):
+            hy_sdk_runtime._acquire_client(
+                profile, api_key=None, data_dir=tmp_path, http_transport_factory=None
+            )
+    else:
+        hy_sdk_runtime._acquire_client(
+            profile, api_key=None, data_dir=tmp_path, http_transport_factory=None
+        )
+
+    assert order == ["client", "confine"]
diff --git a/memhub/apps/core/tests/test_memory_library_all_scenes.py b/memhub/apps/core/tests/test_memory_library_all_scenes.py
index 439aa18..d53dd41 100644
--- a/memhub/apps/core/tests/test_memory_library_all_scenes.py
+++ b/memhub/apps/core/tests/test_memory_library_all_scenes.py
@@ -6,6 +6,7 @@ and delete pass the memory's kind/variant through to the SDK.
 """
 
 import asyncio
+from contextlib import asynccontextmanager
 
 import pytest
 from httpx import ASGITransport, AsyncClient
@@ -55,6 +56,9 @@ def chat_row(memory_id, content, *, source="cursor"):
     }
 
 
+EXPERIENCE_MEMORY_AT = 1_789_990_000
+
+
 def experience_row(memory_id, task, solution, *, variant="working", source="codebuddy"):
     return {
         "experience_id": memory_id,
@@ -63,6 +67,8 @@ def experience_row(memory_id, task, solution, *, variant="working", source="code
         "task": task,
         "solution": solution,
         "source": source,
+        # Conversation time, epoch seconds like chat (hy-memory 2.0.7).
+        "memory_at": EXPERIENCE_MEMORY_AT,
         "created_at": "2026-10-01T08:00:00",
         "updated_at": "2026-10-02T08:00:00",
         "_variant": variant,
@@ -305,12 +311,13 @@ def run_with_sdk(tmp_path, monkeypatch, scenario):
             async def close(self):
                 return None
 
+        @asynccontextmanager
         async def write_client():
-            return sdk
+            yield sdk
 
         async with app.router.lifespan_context(app):
             app.state.sdk_client_handle = Handle()
-            monkeypatch.setattr(app.state.sdk_write_coordinator, "client", write_client)
+            monkeypatch.setattr(app.state.sdk_write_coordinator, "use_client", write_client)
             async with client_for(app) as client:
                 await scenario(app, sdk, client)
 
@@ -446,6 +453,7 @@ def test_keyword_filter_reads_experience_and_fact_text(tmp_path, monkeypatch):
         assert experience["content"] == "Fix flaky tests\nPin the pytest seed"
         assert experience["status"] == "active"
         assert experience["updated_at"] > experience["created_at"]
+        assert experience["memory_at"] == EXPERIENCE_MEMORY_AT
         fact = by_id["fact-1"]
         assert (fact["scene"], fact["kind"], fact["variant"]) == ("coding", "fact", "coding")
         assert fact["content"] == "The pytest suite runs in CI"
@@ -488,6 +496,10 @@ def test_empty_query_lists_all_scenes_with_source_filter_and_totals(tmp_path, mo
             "chat",
             "chat",
         ]
+        assert [item["memory_at"] for item in body["memories"][:2]] == [
+            EXPERIENCE_MEMORY_AT,
+            EXPERIENCE_MEMORY_AT,
+        ]
         call = sdk.list_calls[0]
         assert call["scene"] == "all"
         assert call["include_sources"] is None
@@ -527,6 +539,7 @@ def test_detail_edit_delete_pass_kind_and_variant(tmp_path, monkeypatch):
         assert record["task"] == "Old task"
         assert record["solution"] == "Old solution"
         assert record["content"] == "Old task\nOld solution"
+        assert record["memory_at"] == EXPERIENCE_MEMORY_AT
         assert sdk.get_calls[-1] == ("exp-1", "experience", "working")
 
         # Chat detail with no namespace keeps the plain SDK call.
@@ -692,5 +705,6 @@ def test_home_summary_counts_what_the_memory_page_lists(tmp_path, monkeypatch):
         assert (home_list["offset"], home_list["limit"], home_list["order"]) == (0, 1, "desc")
         assert body["latest"]["memory_id"] == "exp-w"
         assert body["latest"]["scene"] == "working"
+        assert body["latest"]["memory_at"] == EXPERIENCE_MEMORY_AT
 
     run_with_sdk(tmp_path, monkeypatch, scenario)
diff --git a/memhub/apps/core/tests/test_memory_server_durable_queue.py b/memhub/apps/core/tests/test_memory_server_durable_queue.py
index df5d410..062accc 100644
--- a/memhub/apps/core/tests/test_memory_server_durable_queue.py
+++ b/memhub/apps/core/tests/test_memory_server_durable_queue.py
@@ -1187,6 +1187,167 @@ def test_the_drain_tick_writes_a_quiet_live_conversation(tmp_path, monkeypatch)
     assert flushed
 
 
+class _SwitchableProfiles:
+    """Active profile identity that a test can change mid-write."""
+
+    def __init__(self) -> None:
+        self.identity = ("profile-a", 1)
+
+    def active_identity(self):
+        return self.identity
+
+    def snapshot(self, profile_id, revision):
+        return type(
+            "Snapshot", (), {"profile_id": profile_id, "revision": revision, "has_api_key": False}
+        )()
+
+    def resolve_snapshot_secret(self, *_args):
+        return None
+
+
+class _TrackedHandle:
+    def __init__(self, profile, **_kwargs) -> None:
+        self.profile_id = profile.profile_id
+        self.closed = False
+
+    async def client(self):
+        return self
+
+    async def close(self):
+        self.closed = True
+
+
+def _coordinator(monkeypatch, opened: list[_TrackedHandle]):
+    from one_memory import sdk_write_processor
+
+    def handle(profile, **kwargs):
+        value = _TrackedHandle(profile, **kwargs)
+        opened.append(value)
+        return value
+
+    monkeypatch.setattr(sdk_write_processor, "SdkClientHandle", handle)
+    profiles = _SwitchableProfiles()
+    coordinator = sdk_write_processor.SdkWriteCoordinator(
+        profiles=profiles,  # type: ignore[arg-type]
+        secrets=None,  # type: ignore[arg-type]
+        data_dir=None,
+        transport_factory=lambda _profile: None,
+        sedimentation_enabled=lambda: True,
+    )
+    return coordinator, profiles
+
+
+def test_a_profile_change_does_not_close_the_client_of_a_write_in_flight(monkeypatch) -> None:
+    """Build 35 lost a clean install's first write as unknown (AttributeError).
+
+    Saving a model profile changed the active identity while the write ran on
+    the old profile's client. The next coordinator call closed that client at
+    once: the SDK cleared its stores and the SharedRuntime closed mid-`add`.
+    """
+    from one_memory import sdk_write_processor
+
+    opened: list[_TrackedHandle] = []
+    coordinator, profiles = _coordinator(monkeypatch, opened)
+    seen: list[tuple[str, bool]] = []
+    writing = asyncio.Event()
+
+    async def process(self, records, *, request_id, before_hy_invocation):
+        client = await self._client_handle.client()
+        await before_hy_invocation()
+        seen.append((client.profile_id, client.closed))
+        if client.profile_id == "profile-a":
+            await writing.wait()
+        # Still open while its own write runs.
+        seen.append((client.profile_id, client.closed))
+        return sdk_write_processor.SdkWriteReceipt()
+
+    monkeypatch.setattr(sdk_write_processor.SdkWriteProcessor, "process_sdk_job", process)
+
+    async def noop():
+        return None
+
+    async def scenario():
+        first = asyncio.create_task(
+            coordinator.process_sdk_job((), request_id="a", before_hy_invocation=noop)
+        )
+        while not seen:
+            await asyncio.sleep(0)
+        profiles.identity = ("profile-b", 2)
+        # A second job and an inquiry under the new profile.
+        await coordinator.process_sdk_job((), request_id="b", before_hy_invocation=noop)
+        assert [handle.closed for handle in opened] == [False, False]
+        writing.set()
+        await first
+        old, new = opened
+        # Retired after its last use, and new work used the new profile.
+        assert old.closed is True and new.closed is False
+        await coordinator.close()
+        assert new.closed is True
+
+    asyncio.run(scenario())
+    assert seen == [
+        ("profile-a", False),
+        ("profile-b", False),
+        ("profile-b", False),
+        ("profile-a", False),
+    ]
+
+
+def test_an_unused_stale_handle_still_closes_at_once(monkeypatch) -> None:
+    opened: list[_TrackedHandle] = []
+    coordinator, profiles = _coordinator(monkeypatch, opened)
+
+    async def scenario():
+        async with coordinator.use_client() as client:
+            assert client.profile_id == "profile-a"
+        assert opened[0].closed is False
+        profiles.identity = ("profile-b", 2)
+        async with coordinator.use_client() as client:
+            assert client.profile_id == "profile-b"
+            assert opened[0].closed is True
+        assert opened[1].closed is False
+        await coordinator.close()
+        assert opened[1].closed is True
+
+    asyncio.run(scenario())
+
+
+def test_a_client_that_fails_to_construct_parks_the_job_instead_of_unknown(tmp_path) -> None:
+    """Nothing was sent when the client could not be built, so it retries on its own."""
+    from one_memory.sdk_write_processor import SdkWriteProcessor
+
+    class _BrokenHandle:
+        async def client(self):
+            raise AttributeError("'RustBindingsAPI' object has no attribute 'bindings'")
+
+    path = tmp_path / "one-memory.sqlite3"
+    queue = DurableWriteQueue(path)
+    record = queue.submit(_request())
+    _claim_through_drain(queue, record.job_id)
+
+    class _Supervisor:
+        async def process_sdk_job(self, _records, **kwargs):
+            return await SdkWriteProcessor(_BrokenHandle()).process_sdk_job(  # type: ignore[arg-type]
+                (_sdk_evidence("e1"),), **kwargs
+            )
+
+    asyncio.run(
+        execute_job(
+            record.job_id,
+            queue=queue,
+            supervisor=_Supervisor(),
+            evidence=_EvidenceStore({"e1"}),
+        )
+    )
+
+    [row] = _rows(path)
+    assert (row["state"], row["hy_invoked"]) == (WAITING_FOR_MODEL, 0)
+    assert row["last_error"] == "hy_memory_client_unavailable"
+    later = datetime.now(UTC) + timedelta(minutes=5)
+    assert queue.readmit_model_waiters(now=later) == 1
+    assert _state(path, record.job_id) == PENDING
+
+
 def test_sdk_unknown_outcome_inquiry_uses_only_public_receipt() -> None:
     from one_memory.sdk_write_processor import SdkWriteProcessor, SdkWriteReceipt
 
@@ -1699,6 +1860,46 @@ def test_a_raw_provider_message_is_never_stored_as_the_error(tmp_path) -> None:
     assert stored == "RuntimeError"
 
 
+def test_a_failed_job_logs_where_it_failed_but_never_the_message(tmp_path, caplog) -> None:
+    """`error_code=AttributeError` alone named no location; frames, never the message."""
+    path = tmp_path / "one-memory.sqlite3"
+    queue = DurableWriteQueue(path)
+    record = queue.submit(_request())
+    _claim_through_drain(queue, record.job_id)
+
+    def closed_store(client):
+        return client.private_store_name.write
+
+    class _Failing:
+        async def process_sdk_job(self, _records, **kwargs):
+            await kwargs["before_hy_invocation"]()
+            try:
+                closed_store(type("Client", (), {"private_store_name": None})())
+            except AttributeError as error:
+                raise RuntimeError("quoting 'my secret diary'") from error
+
+    with caplog.at_level(logging.ERROR, logger="one_memory.memory_server.durable_queue"):
+        asyncio.run(
+            execute_job(
+                record.job_id,
+                queue=queue,
+                supervisor=_Failing(),
+                evidence=_EvidenceStore({"e1"}),
+            )
+        )
+
+    [line] = [
+        message.getMessage()
+        for message in caplog.records
+        if "durable_write_job_failed" in message.getMessage()
+    ]
+    assert f"job_id={record.job_id} error_code=RuntimeError chain=RuntimeError@" in line
+    assert ";AttributeError@" in line
+    assert "test_memory_server_durable_queue.py:" in line and ":closed_store" in line
+    assert "secret diary" not in line
+    assert "private_store_name" not in line
+
+
 def test_a_machine_error_code_is_preserved(tmp_path) -> None:
     """Recognised codes survive; only free-form text is dropped."""
     from one_memory.memory_server.durable_queue import safe_error_code
diff --git a/memhub/apps/core/tests/test_sdk_client_construction.py b/memhub/apps/core/tests/test_sdk_client_construction.py
new file mode 100644
index 0000000..93f828a
--- /dev/null
+++ b/memhub/apps/core/tests/test_sdk_client_construction.py
@@ -0,0 +1,61 @@
+"""SDK clients of one home are constructed one at a time.
+
+Build 35, clean install: right after the model was activated, the write
+drain's client and a UI read client started in the same second. Each opened
+Chroma on the same empty vector directory, Chroma's shared System cache has
+no lock around that first open, and neither client reached `VectorStore
+ready`. The first write job was lost as unknown.
+"""
+
+import threading
+import time
+
+from one_memory import hy_sdk_runtime
+from one_memory.hy_sdk_runtime import acquire_client, builtin_retrieval_snapshot
+
+
+def test_two_clients_on_a_fresh_home_are_constructed_one_at_a_time(tmp_path, monkeypatch):
+    real = hy_sdk_runtime.HyMemoryClient
+    state = {"inside": 0, "most": 0}
+    count = threading.Lock()
+
+    def recording_client(**options):
+        with count:
+            state["inside"] += 1
+            state["most"] = max(state["most"], state["inside"])
+        try:
+            # Widen the window two unserialized constructors would share.
+            time.sleep(0.2)
+            return real(**options)
+        finally:
+            with count:
+                state["inside"] -= 1
+
+    monkeypatch.setattr(hy_sdk_runtime, "HyMemoryClient", recording_client)
+    start = threading.Barrier(2)
+    leases: list[hy_sdk_runtime.ClientLease] = []
+    failures: list[BaseException] = []
+
+    def open_client():
+        start.wait()
+        try:
+            leases.append(
+                acquire_client(builtin_retrieval_snapshot(), api_key=None, data_dir=tmp_path)
+            )
+        except BaseException as error:  # noqa: BLE001 - reported below
+            failures.append(error)
+
+    threads = [threading.Thread(target=open_client) for _ in range(2)]
+    for thread in threads:
+        thread.start()
+    for thread in threads:
+        thread.join(120)
+    try:
+        assert failures == []
+        assert len(leases) == 2
+        assert state["most"] == 1
+        for lease in leases:
+            assert lease.client.memory_stats(user_id="local") is not None
+    finally:
+        for lease in leases:
+            lease.close()
diff --git a/memhub/apps/core/tests/test_sdk_data_plane_migration.py b/memhub/apps/core/tests/test_sdk_data_plane_migration.py
index b0fb64e..e098376 100644
--- a/memhub/apps/core/tests/test_sdk_data_plane_migration.py
+++ b/memhub/apps/core/tests/test_sdk_data_plane_migration.py
@@ -37,6 +37,7 @@ from one_memory.sdk_migration_runner import (
     abandoned_legacy_work,
     commit_sdk_cutover,
     move_pending_dispatches,
+    prepare_legacy_work_for_cutover,
     prepare_legacy_write_queue_for_cutover,
     recover_abandoned_staging,
     resume_legacy_cleanup,
@@ -137,19 +138,38 @@ def test_post_cutover_metadata_repair_restores_source_and_original_times(store):
     )
     coding = data_dir / "hy-memory" / "governed-v2" / "coding_cv.sqlite3"
     with sqlite3.connect(coding) as database:
-        database.execute(
-            "CREATE TABLE coding_memory_meta ("
-            "experience_id TEXT,user_id TEXT,agent_id TEXT,solution TEXT,"
-            "source TEXT,created_at TEXT,updated_at TEXT)"
+        # A store written by hy-memory 2.0.6, before the conversation time
+        # column: the repair must add it through the SDK and then fill it.
+        database.executescript(
+            """
+            CREATE TABLE coding_memory_meta (
+                experience_id TEXT PRIMARY KEY, user_id TEXT NOT NULL,
+                agent_id TEXT NOT NULL DEFAULT 'default_agent', task TEXT NOT NULL,
+                search_keys TEXT NOT NULL DEFAULT '[]', solution TEXT NOT NULL,
+                boundary_envs TEXT NOT NULL DEFAULT '', workspace_id TEXT,
+                branch TEXT, session_id TEXT, files TEXT NOT NULL DEFAULT '[]',
+                source_fact_ids TEXT NOT NULL DEFAULT '[]',
+                confidence REAL NOT NULL DEFAULT 0.7,
+                source TEXT NOT NULL DEFAULT 'default',
+                extraction_source TEXT NOT NULL DEFAULT 'auto_extract',
+                created_at TEXT NOT NULL, updated_at TEXT NOT NULL
+            );
+            CREATE TABLE working_fact (
+                fact_id TEXT PRIMARY KEY, user_id TEXT NOT NULL,
+                agent_id TEXT NOT NULL DEFAULT 'default_agent', session_id TEXT,
+                workspace_id TEXT, branch TEXT, source TEXT NOT NULL DEFAULT 'default',
+                statement TEXT NOT NULL, files TEXT NOT NULL DEFAULT '[]',
+                subject TEXT NOT NULL DEFAULT '', details TEXT NOT NULL DEFAULT '',
+                fact_link TEXT, derived_experience_ids TEXT NOT NULL DEFAULT '[]',
+                status TEXT NOT NULL DEFAULT 'active',
+                fact_chain_summary TEXT NOT NULL DEFAULT '', created_at TEXT NOT NULL
+            );
+            """
         )
         database.execute(
-            "CREATE TABLE working_fact ("
-            "fact_id TEXT,user_id TEXT,agent_id TEXT,statement TEXT,"
-            "source TEXT,created_at TEXT)"
-        )
-        database.execute(
-            "INSERT INTO coding_memory_meta VALUES "
-            "('sdk-1','local','scope','remember','default','now','now')"
+            "INSERT INTO coding_memory_meta "
+            "(experience_id,user_id,agent_id,task,solution,source,created_at,updated_at) "
+            "VALUES ('sdk-1','local','scope','remember','remember','default','now','now')"
         )
 
     result = repair_sdk_metadata_from_bundle(
@@ -168,10 +188,11 @@ def test_post_cutover_metadata_repair_restores_source_and_original_times(store):
         2026, 10, 18, tzinfo=UTC
     )
     with sqlite3.connect(coding) as database:
-        source, created_at = database.execute(
-            "SELECT source,created_at FROM coding_memory_meta WHERE experience_id='sdk-1'"
+        source, memory_at, created_at = database.execute(
+            "SELECT source,memory_at,created_at FROM coding_memory_meta WHERE experience_id='sdk-1'"
         ).fetchone()
     assert source == "workbuddy"
+    assert datetime.fromtimestamp(memory_at, tz=UTC) == datetime(2026, 10, 17, tzinfo=UTC)
     assert created_at.startswith("2026-10-18T00:00:00")
 
     repeated = repair_sdk_metadata_from_bundle(
@@ -876,6 +897,116 @@ def test_legacy_work_that_may_have_reached_the_sdk_does_not_block_cutover(tmp_pa
     assert store.authority() is MemoryAuthority.SDK
 
 
+def test_pre_upgrade_unknown_writes_are_settled_at_cutover_not_retried(store):
+    """A migrated 1.0.4 home kept 1000+ jobs unknown forever.
+
+    The legacy pipeline invoked them, so the SDK's write-outcome inquiry knows
+    none of them and the automatic reconcile could never settle one.
+    """
+    from one_memory.memory_server.durable_queue import (
+        CANCELLED,
+        PENDING,
+        RUNNING,
+        UNKNOWN,
+    )
+
+    queue = DurableWriteQueue(store.database_path)
+
+    def job(key: str, state: str, *, invoked: bool) -> str:
+        identifier = queue.submit(
+            WriteRequest(
+                internal_session_id="session-1",
+                source_id="cursor",
+                batch_identity=BatchIdentity(
+                    internal_session_id="session-1", first_round=1, last_round=1
+                ),
+                evidence_ids=["gone"],
+                idempotency_key=key,
+            )
+        ).job_id
+        with main_database(store.database_path).write() as database, database:
+            database.execute(
+                "UPDATE memory_write_jobs SET state=?,hy_invoked=?,"
+                "last_error='model_content_outcome_unknown' WHERE job_id=?",
+                (state, int(invoked), identifier),
+            )
+        return identifier
+
+    # Their Evidence is gone too, which must not block the cutover.
+    unknown = [job(f"unknown-{index}", UNKNOWN, invoked=True) for index in range(3)]
+    sent_running = job("sent-running", RUNNING, invoked=True)
+    finished = job("finished", "completed", invoked=True)
+
+    first = prepare_legacy_work_for_cutover(store.database_path)
+    assert first.abandoned == {"unknown_write_jobs": 4}
+    # Resumed after an interruption: nothing changes, the count is the same.
+    again = prepare_legacy_work_for_cutover(store.database_path)
+    assert again.as_document() == first.as_document()
+
+    with main_database(store.database_path).read() as database:
+        rows = dict(
+            (str(row[0]), (str(row[1]), str(row[2]), int(row[3])))
+            for row in database.execute(
+                "SELECT job_id,state,last_error,hy_invoked FROM memory_write_jobs"
+            )
+        )
+    for identifier in (*unknown, sent_running):
+        assert rows[identifier] == (CANCELLED, "abandoned_at_cutover", 1)
+    assert rows[finished][0] == "completed"
+    surviving = DurableWriteQueue(store.database_path, sdk_authority=True)
+    assert surviving.unknown_jobs() == ()
+    # Never requeued: they may already have written.
+    assert surviving.retry_unknown_confirmed(unknown[0]) == 0
+    assert surviving.readmit_parked() == 0
+    assert PENDING not in {state for state, _, _ in rows.values()}
+
+
+def test_retrying_unknown_writes_works_after_the_legacy_tables_are_dropped(store):
+    """The retry button raised `no such table: memory_tasks` after an upgrade."""
+    from one_memory.memory_server.durable_queue import PENDING, UNKNOWN
+    from one_memory.target_schema import drop_legacy_tables
+
+    queue = DurableWriteQueue(store.database_path)
+    identifier = queue.submit(
+        WriteRequest(
+            internal_session_id="session-1",
+            source_id="cursor",
+            batch_identity=BatchIdentity(
+                internal_session_id="session-1", first_round=1, last_round=1
+            ),
+            evidence_ids=["e1"],
+            idempotency_key="after-upgrade",
+        )
+    ).job_id
+    drop_legacy_tables(store.database_path)
+    assert legacy_tables_present(store.database_path) == ()
+    old = datetime(2000, 1, 1, tzinfo=UTC).isoformat()
+
+    def make_unknown(code: str, attempts: int) -> None:
+        with main_database(store.database_path).write() as database, database:
+            database.execute(
+                "UPDATE memory_write_jobs SET state=?,hy_invoked=1,last_error=?,"
+                "attempts=?,updated_at=? WHERE job_id=?",
+                (UNKNOWN, code, attempts, old, identifier),
+            )
+
+    for sdk_authority in (True, False):
+        surviving = DurableWriteQueue(store.database_path, sdk_authority=sdk_authority)
+        make_unknown("model_content_outcome_unknown", 1)
+        assert surviving.resolve_unknown_from_linked_output(identifier) is False
+        assert surviving.retry_unknown_confirmed(identifier) == 1
+        assert surviving.get("after-upgrade").state == PENDING
+
+        make_unknown("hy_memory_timeout", 1)
+        since = datetime.now(UTC)
+        assert surviving.retry_timed_out_unknown_after_reconcile(older_than=since) == 1
+        make_unknown("model_content_outcome_unknown", 1)
+        assert surviving.retry_first_unknown_after_reconcile(older_than=since) == 1
+        if sdk_authority:
+            # The SDK-authority queue never brings a dropped table back.
+            assert legacy_tables_present(store.database_path) == ()
+
+
 def _protective(home: Path) -> Path:
     return home / "hy-memory" / ".governed-v2.pre-migration-0123456789abcdef"
 
diff --git a/memhub/apps/core/tests/test_sdk_migration_bundle.py b/memhub/apps/core/tests/test_sdk_migration_bundle.py
index 19381b4..5f06d89 100644
--- a/memhub/apps/core/tests/test_sdk_migration_bundle.py
+++ b/memhub/apps/core/tests/test_sdk_migration_bundle.py
@@ -91,7 +91,7 @@ def test_bundle_contains_management_sdk_config_and_verified_manifest(tmp_path):
     assert manifest["bundle_id"] == "migration-1"
     assert manifest["sdk_namespace"] == "governed-v2"
     assert manifest["runtime_build"] == "build-13"
-    assert manifest["snapshot_sdk_version"] == "2.0.6"
+    assert manifest["snapshot_sdk_version"] == "2.0.8"
     paths = {item["path"]: item for item in manifest["files"]}
     assert {
         "management/one-memory.sqlite3",
diff --git a/memhub/apps/core/uv.lock b/memhub/apps/core/uv.lock
index 83e11dd..64584a9 100644
--- a/memhub/apps/core/uv.lock
+++ b/memhub/apps/core/uv.lock
@@ -991,7 +991,7 @@ wheels = [
 
 [[package]]
 name = "hy-memory"
-version = "2.0.6"
+version = "2.0.8"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
     { name = "chromadb" },
@@ -1008,9 +1008,9 @@ dependencies = [
     { name = "scikit-learn" },
     { name = "tiktoken" },
 ]
-sdist = { url = "https://files.pythonhosted.org/packages/18/98/312ac8bcbe41e63200a93d296300b2e6748e1a532dad01f009dad57ba408/hy_memory-2.0.6.tar.gz", hash = "sha256:d0f54d54867ee7e722a8dd1a88a00c0576904435f15ae7899fe640840bff74bc", size = 855200, upload-time = "2026-10-27T07:53:10.022Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/21/15/83ca83b649ead238f186fdb33efdc73615fbba279efbbdca3857c1a86beb/hy_memory-2.0.8.tar.gz", hash = "sha256:0fa0ed1a1160b389a3d7c13a9fb6528b140d7ec23cf49e65f1be5974820a576b", size = 856025, upload-time = "2026-10-28T07:35:47.783Z" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/27/f1/d4866cffe426d77b857168cc694b15df866f3d9f45152c17c148e1595db4/hy_memory-2.0.6-py3-none-any.whl", hash = "sha256:b0e666f8e89d1763bb31c0e67502df74434134a638ff6fd1912de9e936be2483", size = 918955, upload-time = "2026-10-27T07:53:07.686Z" },
+    { url = "https://files.pythonhosted.org/packages/11/6a/c699cac0ff06bd089fceeff90446130763d8187e649dbdb0313014491a10/hy_memory-2.0.8-py3-none-any.whl", hash = "sha256:68bd789df24e51e64bb231e97a8d24ee89b4aa5f9ba14c3e0824589e9213bf7c", size = 921119, upload-time = "2026-10-28T07:35:45.571Z" },
 ]
 
 [[package]]
@@ -1837,7 +1837,7 @@ requires-dist = [
     { name = "fastapi", specifier = ">=0.141.1" },
     { name = "filelock", specifier = ">=3.32.4" },
     { name = "httpx", specifier = ">=0.28.1" },
-    { name = "hy-memory", specifier = "==2.0.6" },
+    { name = "hy-memory", specifier = "==2.0.8" },
     { name = "keyring", specifier = ">=25.7.0" },
     { name = "numpy", specifier = "==2.5.2" },
     { name = "onnxruntime", specifier = "==1.23.2" },
diff --git a/memhub/apps/desktop/src/diagnostics-ipc.test.ts b/memhub/apps/desktop/src/diagnostics-ipc.test.ts
index 3364a75..c84b856 100644
--- a/memhub/apps/desktop/src/diagnostics-ipc.test.ts
+++ b/memhub/apps/desktop/src/diagnostics-ipc.test.ts
@@ -230,39 +230,21 @@ it("filters raw Core INFO and DEBUG while preserving warnings and errors", async
   ).toBe(true);
 });
 
-it("shows HY pipeline metrics without exposing model content", async () => {
-  const dataDir = await mkdtemp(
-    path.join(os.tmpdir(), "om-debug-hy-pipeline-"),
-  );
+it("does not read HY's own log directory", async () => {
+  // MemHub turns the HY pipeline log off; files left by older builds and
+  // HY's text log stay on disk but are not part of the Debug panel.
+  const dataDir = await mkdtemp(path.join(os.tmpdir(), "om-debug-hy-logs-"));
   roots.push(dataDir);
-  const pipelineDir = path.join(
-    dataDir,
-    "hy-memory",
-    "governed-v2",
-    "logs",
-    "default",
-  );
-  await mkdir(pipelineDir, { recursive: true });
-  const target = path.join(pipelineDir, "2026-10-14.log");
-  await writeFile(target, "");
+  const hyLogs = path.join(dataDir, "hy-memory", "governed-v2", "logs");
+  await mkdir(path.join(hyLogs, "default"), { recursive: true });
   const cursor = await captureDebugLogCursor(dataDir);
-  const pipelineEntry = {
-    request_id: "request-safe-1",
-    user_id: "private-user",
-    agent_id: "private-agent",
-    session_id: "private-session",
-    step: "EXTRACT",
-    prompt: "private prompt",
-    content: "private response",
-    parsed: "private parsed output",
-    tokens: 42,
-    elapsed_ms: 123.5,
-    timestamp: "2026-10-14T10:01:00.000",
-  };
-  await appendFile(target, `${JSON.stringify(pipelineEntry)}\n`);
   await writeFile(
-    path.join(dataDir, "hy-memory", "governed-v2", "logs", "hy_memory.log"),
-    `${JSON.stringify({ ...pipelineEntry, request_id: "must-not-appear" })}\n`,
+    path.join(hyLogs, "default", "2026-10-14.log"),
+    `${JSON.stringify({ request_id: "old-pipeline", step: "EXTRACT" })}\n`,
+  );
+  await writeFile(
+    path.join(hyLogs, "hy_memory.log"),
+    "2026-10-14 10:01:00 [WARNING] [-] hy_memory.client: hy text log\n",
   );
 
   const snapshot = await readDebugLogSnapshot(
@@ -271,29 +253,11 @@ it("shows HY pipeline metrics without exposing model content", async () => {
     cursor,
   );
 
-  expect(snapshot.entries).toEqual([
-    expect.objectContaining({
-      service: "hy-memory",
-      component: "hy.pipeline",
-      event: "hy.pipeline.extract",
-      message: "EXTRACT · 42 tokens · 124 ms",
-      duration_ms: 123.5,
-      trace_id: "request-safe-1",
-      details: { phase: "EXTRACT", count: 42 },
-      source: "pipeline",
-    }),
-  ]);
-  const serialized = JSON.stringify(snapshot);
-  for (const secret of [
-    "private prompt",
-    "private response",
-    "private parsed output",
-    "private-user",
-    "private-agent",
-    "private-session",
-  ]) {
-    expect(serialized).not.toContain(secret);
-  }
+  expect(snapshot.entries).toEqual([]);
+  expect(cursor).toEqual({
+    diagnosticLogs: [],
+    serviceLogs: expect.any(Array),
+  });
 });
 
 it("logs a renderer ui-action with action, page, phase and the shared trace id", async () => {
diff --git a/memhub/apps/desktop/src/diagnostics-ipc.ts b/memhub/apps/desktop/src/diagnostics-ipc.ts
index e84295b..a11ca9b 100644
--- a/memhub/apps/desktop/src/diagnostics-ipc.ts
+++ b/memhub/apps/desktop/src/diagnostics-ipc.ts
@@ -18,7 +18,6 @@ const UI_TRACE = "one-memory:log-ui-trace";
 const DEBUG_LOGS = "one-memory:get-debug-logs";
 const DEBUG_LOG_NAME =
   /^diagnostic-(runtime|core|host-bridge|desktop|cli)\.jsonl(?:\.\d+)?$/;
-const PIPELINE_LOG_NAME = /^\d{4}-\d{2}-\d{2}\.log$/;
 const SERVICE_LOG_NAMES = [
   "service.log",
   "service-error.log",
@@ -30,24 +29,16 @@ interface ServiceLogCursor {
   files: Array<{ device: number; inode: number; offset: number }>;
 }
 
-interface PipelineLogCursor {
-  relativePath: string;
-  device: number;
-  inode: number;
-  offset: number;
-}
-
 export interface DebugLogCursor {
   diagnosticLogs: Array<{ device: number; inode: number; offset: number }>;
   serviceLogs: ServiceLogCursor[];
-  pipelineLogs: PipelineLogCursor[];
 }
 
 export interface DebugLogEntry {
   id: string;
   timestamp: string;
   level: "debug" | "info" | "warn" | "error";
-  service: "runtime" | "core" | "host-bridge" | "desktop" | "cli" | "hy-memory";
+  service: "runtime" | "core" | "host-bridge" | "desktop" | "cli";
   component: string;
   event: string;
   outcome: string;
@@ -57,7 +48,7 @@ export interface DebugLogEntry {
   trace_id?: string;
   details?: Record<string, string | number | boolean | null>;
   message?: string;
-  source: "diagnostic" | "stdout" | "stderr" | "pipeline";
+  source: "diagnostic" | "stdout" | "stderr";
 }
 
 export interface DebugLogSnapshot {
@@ -259,7 +250,6 @@ export async function readDebugLogSnapshot(
   cursor: DebugLogCursor = {
     diagnosticLogs: [],
     serviceLogs: [],
-    pipelineLogs: [],
   },
 ): Promise<DebugLogSnapshot> {
   const started = Number.isFinite(startedAt) ? startedAt : Date.now();
@@ -300,7 +290,6 @@ export async function readDebugLogSnapshot(
       );
     }
   }
-  entries.push(...(await readPipelineLogEntries(dataDir, cursor, started)));
   entries.sort((left, right) =>
     left.timestamp === right.timestamp
       ? left.id.localeCompare(right.id)
@@ -319,7 +308,6 @@ export async function captureDebugLogCursor(
 ): Promise<DebugLogCursor> {
   const directory = path.join(dataDir, "logs");
   const names = await readdir(directory).catch(() => []);
-  const pipelineRoot = path.join(dataDir, "hy-memory", "governed-v2", "logs");
   return {
     diagnosticLogs: (
       await Promise.all(
@@ -355,7 +343,6 @@ export async function captureDebugLogCursor(
         };
       }),
     ),
-    pipelineLogs: await pipelineLogFiles(pipelineRoot),
   };
 }
 
@@ -462,8 +449,8 @@ async function readServiceLogEntries(
       if (!message) continue;
       const level = serviceLogLevel(message, cursor.name);
       // Raw stdout is an implementation trace, not a product event. Core's
-      // structured diagnostics and HY pipeline logs already carry meaningful
-      // successful work with request/task identity. Showing raw INFO/DEBUG
+      // structured diagnostics already carry meaningful successful work with
+      // request/task identity. Showing raw INFO/DEBUG
       // here duplicated those events and turned polling access logs, file
       // watchers and library initialization into thousands of fake "useful"
       // rows. Keep raw output only when it needs attention.
@@ -543,135 +530,6 @@ function logLineTimestamp(
   return new Date(Math.max(startedAt, fallback)).toISOString();
 }
 
-async function pipelineLogFiles(root: string): Promise<PipelineLogCursor[]> {
-  const rootInfo = await lstat(root).catch(() => null);
-  if (!rootInfo?.isDirectory() || rootInfo.isSymbolicLink()) return [];
-  const files: PipelineLogCursor[] = [];
-  const visit = async (directory: string): Promise<void> => {
-    for (const entry of await readdir(directory, { withFileTypes: true })) {
-      const target = path.join(directory, entry.name);
-      if (entry.isSymbolicLink()) continue;
-      if (entry.isDirectory()) {
-        await visit(target);
-        continue;
-      }
-      // The same root also contains HY's human-readable hy_memory.log.
-      // Pipeline files are date-named JSONL; accepting every .log file makes
-      // the Debug poller repeatedly parse unrelated text and still display no
-      // HY entries.
-      if (!entry.isFile() || !PIPELINE_LOG_NAME.test(entry.name)) continue;
-      const info = await safeRegularFile(target);
-      if (!info) continue;
-      files.push({
-        relativePath: path.relative(root, target),
-        device: info.dev,
-        inode: info.ino,
-        offset: info.size,
-      });
-    }
-  };
-  await visit(root);
-  return files;
-}
-
-async function readPipelineLogEntries(
-  dataDir: string,
-  cursor: DebugLogCursor,
-  startedAt: number,
-): Promise<DebugLogEntry[]> {
-  const root = path.join(dataDir, "hy-memory", "governed-v2", "logs");
-  const current = await pipelineLogFiles(root);
-  const baseline = new Map(
-    cursor.pipelineLogs.map((file) => [
-      `${file.device}:${file.inode}`,
-      file.offset,
-    ]),
-  );
-  const entries: DebugLogEntry[] = [];
-  for (const file of current) {
-    const target = path.join(root, file.relativePath);
-    const offset = baseline.get(`${file.device}:${file.inode}`) ?? 0;
-    const info = await safeRegularFile(target);
-    if (!info || info.size <= offset) continue;
-    const content = await readFileFromOffset(target, offset).catch(() => "");
-    let lineNumber = 0;
-    for (const line of content.split("\n")) {
-      lineNumber += 1;
-      if (!line.trim()) continue;
-      const entry = pipelineLogEntry(
-        line,
-        file,
-        lineNumber,
-        offset,
-        startedAt,
-        info.mtimeMs,
-      );
-      if (entry) entries.push(entry);
-    }
-  }
-  return entries;
-}
-
-function pipelineLogEntry(
-  line: string,
-  file: PipelineLogCursor,
-  lineNumber: number,
-  offset: number,
-  startedAt: number,
-  fallbackTime: number,
-): DebugLogEntry | null {
-  let value: Record<string, unknown>;
-  try {
-    value = JSON.parse(line) as Record<string, unknown>;
-  } catch {
-    return null;
-  }
-  const step = typeof value.step === "string" ? value.step.trim() : "";
-  if (!step || !/^[A-Za-z0-9_-]{1,80}$/.test(step)) return null;
-  const rawTimestamp =
-    typeof value.timestamp === "string" ? value.timestamp : "";
-  const parsedTimestamp = Date.parse(rawTimestamp);
-  const timestamp = new Date(
-    Math.max(
-      startedAt,
-      Number.isFinite(parsedTimestamp) ? parsedTimestamp : fallbackTime,
-    ),
-  ).toISOString();
-  const elapsed = safeFiniteNumber(value.elapsed_ms);
-  const tokens = safeFiniteNumber(value.tokens);
-  const requestId =
-    typeof value.request_id === "string" && value.request_id.length <= 256
-      ? value.request_id
-      : undefined;
-  return {
-    id: `hy-pipeline:${file.device}:${file.inode}:${offset}:${lineNumber}`,
-    timestamp,
-    level: "info",
-    service: "hy-memory",
-    component: "hy.pipeline",
-    event: `hy.pipeline.${step.toLowerCase().replaceAll("_", "-")}`,
-    outcome: "observed",
-    ...(elapsed === undefined ? {} : { duration_ms: elapsed }),
-    ...(requestId ? { trace_id: requestId } : {}),
-    details: {
-      phase: step,
-      ...(tokens === undefined ? {} : { count: tokens }),
-    },
-    message: [
-      step,
-      ...(tokens === undefined ? [] : [`${Math.round(tokens)} tokens`]),
-      ...(elapsed === undefined ? [] : [`${Math.round(elapsed)} ms`]),
-    ].join(" · "),
-    source: "pipeline",
-  };
-}
-
-function safeFiniteNumber(value: unknown): number | undefined {
-  return typeof value === "number" && Number.isFinite(value)
-    ? value
-    : undefined;
-}
-
 export function installDiagnosticsIpc(
   options: DiagnosticIpcOptions,
 ): () => void {
diff --git a/memhub/apps/desktop/src/host-control-ownership.test.ts b/memhub/apps/desktop/src/host-control-ownership.test.ts
index 67b6be5..993fd5f 100644
--- a/memhub/apps/desktop/src/host-control-ownership.test.ts
+++ b/memhub/apps/desktop/src/host-control-ownership.test.ts
@@ -15,6 +15,7 @@ const fake = vi.hoisted(() => ({
   missingBroker: false,
   hostCapability: true,
   open: vi.fn(),
+  refreshHosts: vi.fn(async (_layouts: unknown, _options: unknown) => ({})),
   host: vi.fn<
     (input: {
       hostId: string;
@@ -274,7 +275,9 @@ beforeEach(async () => {
   fake.open.mockResolvedValue({
     hostGrantStore: {},
     repairInstalledCommand: vi.fn(async () => []),
+    refreshInstalledHosts: fake.refreshHosts,
   });
+  fake.refreshHosts.mockClear();
   fake.host.mockResolvedValue({ state: "synthetic-remote-result" });
   Object.defineProperty(process, "resourcesPath", {
     configurable: true,
@@ -590,5 +593,11 @@ describe("packaged Desktop host-control ownership", () => {
     expect(fake.ensureRuntime).not.toHaveBeenCalled();
     expect(fake.createClient).not.toHaveBeenCalled();
     expect(fake.windows).toHaveLength(1);
+    // The embedded Runtime is this process, so Desktop refreshes the hosts
+    // it installed once the Runtime is up, as a standalone one would itself.
+    expect(fake.refreshHosts).toHaveBeenCalledTimes(1);
+    expect(fake.refreshHosts.mock.calls[0]?.[1]).toMatchObject({
+      diagnostics: expect.anything(),
+    });
   });
 });
diff --git a/memhub/apps/desktop/src/host-permissions.test.ts b/memhub/apps/desktop/src/host-permissions.test.ts
index 6593032..f2be754 100644
--- a/memhub/apps/desktop/src/host-permissions.test.ts
+++ b/memhub/apps/desktop/src/host-permissions.test.ts
@@ -21,6 +21,8 @@ import {
   claudeCodeLayout,
   codexLayout,
   cursorLayout,
+  MCP_TOOLSET_KEY,
+  MCP_TOOLSET_VERSION,
   mcpConfigPath,
   piLayout,
   planRecallInstall,
@@ -29,6 +31,7 @@ import {
   workbuddyLayout,
   workbuddyScatteredLayout,
 } from "@one-memory/host-bridge/installer";
+import type { DiagnosticInput } from "@one-memory/runtime-control";
 import { afterEach, describe, expect, it, vi } from "vitest";
 import { probeRecallInstall } from "../../../packages/host-bridge/src/recall-probe.js";
 import { skillFilePath } from "../../../packages/host-bridge/src/skill-drift.js";
@@ -1384,3 +1387,197 @@ describe("upgrading installs left by an earlier release", () => {
     );
   });
 });
+
+describe("refreshing installed hosts after an upgrade", () => {
+  function recorder() {
+    const events: DiagnosticInput[] = [];
+    return {
+      events,
+      sink: {
+        log: async (event: DiagnosticInput) => {
+          events.push(event);
+        },
+      },
+    };
+  }
+
+  async function linkedWorkbuddy(context: Awaited<ReturnType<typeof fixture>>) {
+    const layout = workbuddyLayout(path.join(context.root, "workbuddy"));
+    const preview = await context.permissions.previewInstall(layout);
+    const result = await context.permissions.install(
+      layout,
+      preview.fingerprint,
+    );
+    await context.permissions.setPermission(
+      layout,
+      approval(required(result.permission.receipt)),
+    );
+    return layout;
+  }
+
+  /** Put back the skill an earlier release wrote, journaled as ours. */
+  async function ageSkill(layout: ReturnType<typeof workbuddyLayout>) {
+    const skillPath = required(skillFilePath(layout));
+    const aged = (await readFile(skillPath, "utf8"))
+      .replace(/skill-version=\d+/, "skill-version=15")
+      .replace("`MemHub_recall`", "`MemHub_ask`");
+    await writeFile(skillPath, aged);
+    const journal = JSON.parse(await readFile(layout.journalPath, "utf8"));
+    journal.installed_skill_hash = createHash("sha256")
+      .update(aged)
+      .digest("hex");
+    await writeFile(layout.journalPath, JSON.stringify(journal));
+    return aged;
+  }
+
+  it("rewrites an old skill on the first start of a build, and only once per build", async () => {
+    const context = await fixture();
+    const layout = await linkedWorkbuddy(context);
+    const current = await readFile(required(skillFilePath(layout)), "utf8");
+    await ageSkill(layout);
+    const { events, sink } = recorder();
+
+    const first = await context.permissions.refreshInstalledHosts([layout], {
+      build: "34",
+      diagnostics: sink,
+    });
+    expect(first).toMatchObject({
+      state: "refreshed",
+      synced: ["workbuddy"],
+      failed: [],
+    });
+    expect(await readFile(required(skillFilePath(layout)), "utf8")).toBe(
+      current,
+    );
+
+    // Same build again: one file read, nothing rewritten, even if stale.
+    const aged = await ageSkill(layout);
+    expect(
+      await context.permissions.refreshInstalledHosts([layout], {
+        build: "34",
+        diagnostics: sink,
+      }),
+    ).toMatchObject({ state: "already_refreshed", synced: [] });
+    expect(await readFile(required(skillFilePath(layout)), "utf8")).toBe(aged);
+
+    // The next build refreshes again.
+    expect(
+      await context.permissions.refreshInstalledHosts([layout], {
+        build: "35",
+      }),
+    ).toMatchObject({ state: "refreshed", synced: ["workbuddy"] });
+    expect(await readFile(required(skillFilePath(layout)), "utf8")).toBe(
+      current,
+    );
+
+    // Diagnostics carry host ids and counts only, never file content.
+    expect(events.map((event) => event.event)).toEqual([
+      "host_refresh.host_updated",
+      "host_refresh.completed",
+      "host_refresh.skipped",
+    ]);
+    expect(events[0].details).toEqual({ source_id: "workbuddy" });
+    expect(events[1].details).toEqual({
+      count: 1,
+      updated_count: 1,
+      error_count: 0,
+    });
+    expect(JSON.stringify(events)).not.toContain("MemHub_");
+    expect(JSON.stringify(events)).not.toContain(context.root);
+  });
+
+  it("leaves a skill the user edited untouched", async () => {
+    const context = await fixture();
+    const layout = await linkedWorkbuddy(context);
+    const edited = `${await ageSkill(layout)}\nUser addition\n`;
+    await writeFile(required(skillFilePath(layout)), edited);
+
+    expect(
+      await context.permissions.refreshInstalledHosts([layout], {
+        build: "34",
+      }),
+    ).toMatchObject({ state: "refreshed", synced: [], failed: [] });
+    expect(await readFile(required(skillFilePath(layout)), "utf8")).toBe(
+      edited,
+    );
+  });
+
+  it("rewrites an MCP entry without the tool set stamp, in the plugin manifest too", async () => {
+    const context = await fixture();
+    const layout = await linkedWorkbuddy(context);
+    const plugin = required(layout.plugin);
+    const manifests = [
+      path.join(plugin.installRoot, ".codebuddy-plugin", "plugin.json"),
+      path.join(
+        plugin.marketplaceRoot,
+        "plugins",
+        plugin.pluginName,
+        ".codebuddy-plugin",
+        "plugin.json",
+      ),
+    ];
+    // What an earlier release wrote: the same entry with no stamp, journaled.
+    const journal = JSON.parse(await readFile(layout.journalPath, "utf8"));
+    for (const manifestPath of manifests) {
+      const manifest = JSON.parse(await readFile(manifestPath, "utf8"));
+      expect(manifest.mcpServers["one-memory"].env[MCP_TOOLSET_KEY]).toBe(
+        MCP_TOOLSET_VERSION,
+      );
+      delete manifest.mcpServers["one-memory"].env[MCP_TOOLSET_KEY];
+      const text = `${JSON.stringify(manifest, null, 2)}\n`;
+      await writeFile(manifestPath, text);
+      const entry = journal.plugin_files.find(
+        (file: { path: string }) => file.path === manifestPath,
+      );
+      entry.hash = createHash("sha256").update(text).digest("hex");
+      journal.installed_server_hash = createHash("sha256")
+        .update(JSON.stringify(manifest.mcpServers["one-memory"]))
+        .digest("hex");
+    }
+    await writeFile(layout.journalPath, JSON.stringify(journal));
+    const granted = await context.permissions.getPermission(layout);
+    expect(granted.state).toBe("approved");
+    expect(
+      (await readInstalledRecallCommand(layout))?.env[MCP_TOOLSET_KEY],
+    ).toBeUndefined();
+
+    expect(
+      await context.permissions.refreshInstalledHosts([layout], {
+        build: "34",
+      }),
+    ).toMatchObject({ synced: ["workbuddy"], failed: [] });
+    for (const manifestPath of manifests) {
+      const manifest = JSON.parse(await readFile(manifestPath, "utf8"));
+      expect(manifest.mcpServers["one-memory"].env[MCP_TOOLSET_KEY]).toBe(
+        MCP_TOOLSET_VERSION,
+      );
+    }
+    // Rewriting our own entry keeps the user's authorization.
+    expect(await context.permissions.getPermission(layout)).toEqual(granted);
+  });
+
+  it("rewrites a config-file MCP entry without the stamp", async () => {
+    const context = await fixture();
+    await install(context);
+    const configPath = mcpConfigPath(context.layout);
+    const config = JSON.parse(await readFile(configPath, "utf8"));
+    delete config.mcpServers["one-memory"].env[MCP_TOOLSET_KEY];
+    await writeFile(configPath, `${JSON.stringify(config, null, 2)}\n`);
+    const journal = JSON.parse(
+      await readFile(context.layout.journalPath, "utf8"),
+    );
+    journal.installed_server_hash = createHash("sha256")
+      .update(JSON.stringify(config.mcpServers["one-memory"]))
+      .digest("hex");
+    await writeFile(context.layout.journalPath, JSON.stringify(journal));
+
+    expect(
+      await context.permissions.refreshInstalledHosts([context.layout]),
+    ).toMatchObject({ state: "refreshed", synced: ["cursor"], failed: [] });
+    const refreshed = JSON.parse(await readFile(configPath, "utf8"));
+    expect(refreshed.mcpServers["one-memory"].env[MCP_TOOLSET_KEY]).toBe(
+      MCP_TOOLSET_VERSION,
+    );
+    expect(refreshed.mcpServers.neighbor).toEqual({ command: "keep" });
+  });
+});
diff --git a/memhub/apps/desktop/src/main.ts b/memhub/apps/desktop/src/main.ts
index 48b3dfe..791a82a 100644
--- a/memhub/apps/desktop/src/main.ts
+++ b/memhub/apps/desktop/src/main.ts
@@ -1369,6 +1369,24 @@ async function initialize(isQuitting: () => boolean): Promise<void> {
     });
     await runtimeSupervisor.start();
     if (isQuitting()) return;
+    // A standalone Runtime refreshes the hosts it installed as it starts, so
+    // an upgrade through the App or `memhub upgrade` is covered there. An
+    // embedded one is this process: do the same here, now it is up. Only a
+    // packaged build, so a source checkout never repoints real hosts at itself.
+    if (localPermissions && app.isPackaged) {
+      void localPermissions
+        .refreshInstalledHosts(
+          hosts.map((host) => host.layout),
+          {
+            build:
+              packagedBuildNumber === undefined
+                ? undefined
+                : String(packagedBuildNumber),
+            diagnostics: hostDiagnostics,
+          },
+        )
+        .catch(() => undefined);
+    }
     if (
       app.isPackaged &&
       process.platform === "darwin" &&
diff --git a/memhub/apps/web/src/DebugLogPanel.tsx b/memhub/apps/web/src/DebugLogPanel.tsx
index 6f6b904..5173afa 100644
--- a/memhub/apps/web/src/DebugLogPanel.tsx
+++ b/memhub/apps/web/src/DebugLogPanel.tsx
@@ -50,7 +50,6 @@ const SERVICE_LABELS: Record<DebugLogEntry["service"], string> = {
   core: "Core",
   "host-bridge": "Host Bridge",
   cli: "CLI",
-  "hy-memory": "HY Memory",
 };
 
 export function DebugLogPanel() {
diff --git a/memhub/apps/web/src/MemoryMap.test.tsx b/memhub/apps/web/src/MemoryMap.test.tsx
index 6cd12a3..5feaf65 100644
--- a/memhub/apps/web/src/MemoryMap.test.tsx
+++ b/memhub/apps/web/src/MemoryMap.test.tsx
@@ -5,7 +5,6 @@ import type {
   MemoryGraphNode,
   MemoryGraphResponse,
   MemoryProcessingSummary,
-  SourceImportReceipt as Receipt,
   SourceResponse,
 } from "@one-memory/contracts";
 import {
@@ -40,12 +39,6 @@ import { afterEach, expect, it, vi } from "vitest";
 import { MemoryMap } from "./MemoryMap";
 import type { MemoryArrival } from "./memory-arrival-state";
 
-const { sourceReceipt } = vi.hoisted(() => ({
-  sourceReceipt: vi.fn(),
-}));
-vi.mock("./api", () => ({
-  getSourceImportReceipt: sourceReceipt,
-}));
 vi.mock("./MemoryGraph3D", () => ({
   MemoryGraph3D: ({
     graph,
@@ -680,7 +673,7 @@ it("D96: the memory page has no raw-dialogue tab and never loads the dialogue li
 it("D96: lists a memory by when its conversation happened, not when it was generated", () => {
   const today = new Date().toISOString();
   const thisYear = new Date().getFullYear();
-  const items = ["第一条", "第二条", "第三条"].map((content, i) => ({
+  const items = ["第一条", "第二条", "第三条", "第四条"].map((content, i) => ({
     ...memories[0],
     memory_id: `t${i}`,
     content,
@@ -695,10 +688,11 @@ it("D96: lists a memory by when its conversation happened, not when it was gener
         total: 519,
         offset: 100,
         limit: 100,
-        source_ids_by_memory: { t0: [], t1: [], t2: [] },
+        source_ids_by_memory: { t0: [], t1: [], t2: [], t3: [] },
         source_time_by_memory: {
           t0: { at: `${thisYear}-10-02T10:54:00`, basis: "conversation" },
           t1: { at: `${thisYear}-09-30T08:00:00`, basis: "import" },
+          t3: { at: `${thisYear}-09-29T08:00:00`, basis: "write" },
         },
       })}
     />,
@@ -710,6 +704,10 @@ it("D96: lists a memory by when its conversation happened, not when it was gener
   expect(
     within(list).getByRole("button", { name: /第二条/ }).textContent,
   ).toContain("导入于 9月30日");
+  // An experience stored without a conversation time names its write time.
+  expect(
+    within(list).getByRole("button", { name: /第四条/ }).textContent,
+  ).toContain("写入于 9月29日");
   // D99: the generated day keeps its own line, so a memory made today from an
   // old conversation still reads as new rather than looking misplaced.
   expect(
@@ -1145,35 +1143,6 @@ const importedSources = [
     imported_conversations: 1,
   },
 ] as SourceResponse[];
-const importedReceipt: Receipt = {
-  source_id: "codex",
-  measured_at: "2026-09-14T00:00:00Z",
-  messages: 5,
-  conversations: 2,
-  person_messages: 2,
-  agent_messages: 2,
-  tool_messages: 0,
-  unknown_author_messages: 1,
-  unknown_time_messages: 1,
-  examples: [
-    {
-      evidence_id: "raw-codex",
-      role: "user",
-      written_by: "person",
-      created_at: null,
-      excerpt: "合成来源原文，不是治理记忆。",
-      truncated: false,
-    },
-  ],
-};
-async function openSourceReceipt() {
-  await userEvent.click(
-    screen.getByRole("button", { name: "图中选择 Codex · 2 段" }),
-  );
-  await userEvent.click(screen.getByText("查看已读内容"));
-  await waitFor(() => expect(sourceReceipt).toHaveBeenCalledWith("codex"));
-}
-
 it("REV09: selecting an imported source retains the canvas and only its authorized node", async () => {
   render(<Host items={[]} sources={importedSources} />);
   const canvas = screen.getByTestId("canvas");
@@ -1188,7 +1157,8 @@ it("REV09: selecting an imported source retains the canvas and only its authoriz
   expect(
     screen.getByRole("list", { name: "地图记忆列表" }).children,
   ).toHaveLength(0);
-  expect(sourceReceipt).not.toHaveBeenCalled();
+  expect(screen.getByRole("heading", { name: "记忆详情" })).toBeTruthy();
+  expect(screen.getByText("还没有已生成的记忆")).toBeTruthy();
   await userEvent.click(screen.getByRole("button", { name: /全部记忆/ }));
   expect(within(canvas).getAllByRole("button")).toHaveLength(2);
   expect(canvas.dataset.selected ?? "").toBe("");
@@ -1210,83 +1180,248 @@ it("REV09: fallback search matches source names and IDs without bypassing the se
   );
   await userEvent.type(search, "cOdEx");
   expect(screen.getByTestId("canvas").dataset.selected).toBe("source:codex");
-  expect(screen.getByText("查看已读内容")).toBeTruthy();
+  expect(screen.getByText("还没有已生成的记忆")).toBeTruthy();
   await userEvent.clear(search);
   await userEvent.type(search, "Cursor");
   expect(screen.queryByTestId("canvas")).toBeNull();
-  expect(screen.queryByText("查看已读内容")).toBeNull();
+  expect(screen.queryByText("正在读取记忆详情…")).toBeNull();
   await userEvent.clear(search);
   expect(screen.getByTestId("canvas").dataset.selected ?? "").toBe("");
-  expect(screen.queryByText("查看已读内容")).toBeNull();
+  expect(screen.queryByText("正在读取记忆详情…")).toBeNull();
 });
 
-it("REV09: source selection uses the actual permission-checked receipt API", async () => {
-  sourceReceipt.mockResolvedValue(importedReceipt);
-  render(<Host items={[]} sources={importedSources} />);
-  await openSourceReceipt();
-  expect(
-    await screen.findByText(importedReceipt.examples[0].excerpt),
-  ).toBeTruthy();
-  expect(screen.getByText("用户 · 时间未知")).toBeTruthy();
-  expect(screen.getByText("原始消息").nextElementSibling?.textContent).toBe(
-    "5",
+it("selecting a source node waits for its settled page, then opens the first memory", async () => {
+  const onQuery = vi.fn();
+  const onSource = vi.fn();
+  const resetPage = vi.fn();
+  const oldPage = {
+    items: memories,
+    total: 2,
+    offset: 0,
+    limit: 20,
+    source_ids_by_memory: { m0: ["codex"], m1: ["codex"] },
+  };
+  const view = render(
+    <Host
+      simplified
+      sources={importedSources}
+      libraryView={{
+        ...pagedView(oldPage),
+        query: "旧查询",
+        settledQuery: "旧查询",
+        onQuery,
+        onSource,
+        resetPage,
+      }}
+    />,
   );
-  // No raw conversation is read by id on the SDK data plane.
-  expect(screen.queryByRole("button", { name: "查看完整对话" })).toBeNull();
+  await act(async () => {});
+  await userEvent.click(screen.getByRole("button", { name: "图中选择 Codex" }));
+  expect(onQuery).toHaveBeenCalledWith("");
+  expect(onSource).toHaveBeenCalledWith("codex");
+  expect(resetPage).toHaveBeenCalledOnce();
+
+  view.rerender(
+    <Host
+      simplified
+      sources={importedSources}
+      libraryView={{
+        ...pagedView(oldPage),
+        source: "codex",
+        settledSource: "",
+        loading: true,
+        onQuery,
+        onSource,
+        resetPage,
+      }}
+    />,
+  );
+  expect(
+    screen
+      .getAllByRole("status")
+      .some((node) => node.textContent === "正在读取记忆详情…"),
+  ).toBe(true);
+
+  const settledItems = [memories[1], memories[0]];
+  view.rerender(
+    <Host
+      simplified
+      items={settledItems}
+      sources={importedSources}
+      libraryView={{
+        ...pagedView({
+          ...oldPage,
+          items: settledItems,
+          source_ids_by_memory: { m1: ["codex"], m0: ["codex"] },
+        }),
+        source: "codex",
+        settledSource: "codex",
+        settledQuery: "",
+        onQuery,
+        onSource,
+        resetPage,
+      }}
+    />,
+  );
+  expect(await screen.findByRole("heading", { name: "第二条" })).toBeTruthy();
+  expect(screen.queryByText("查看已读内容")).toBeNull();
+  expect(document.querySelector(".source-import-receipt")).toBeNull();
 });
 
-it.each(["revoked", "removed", "revoked-hidden"])(
-  "REV09: %s source drops fallback, preview and a late receipt without reopening on reauthorization",
-  async (state) => {
-    let finish: (value: Receipt) => void = () => {};
-    sourceReceipt.mockImplementation(
-      () =>
-        new Promise<Receipt>((resolve) => {
-          finish = resolve;
-        }),
-    );
-    const view = render(<Host items={[]} sources={[importedSources[0]]} />);
-    await openSourceReceipt();
-    view.rerender(
-      <Host
-        items={[]}
-        hidden={state === "revoked-hidden"}
-        sources={
-          state === "removed"
-            ? []
-            : [{ ...importedSources[0], authorized: false }]
-        }
-      />,
-    );
-    expect(screen.queryByTestId("canvas")).toBeNull();
-    expect(document.querySelector(".source-import-receipt")).toBeNull();
-    await act(async () => finish(importedReceipt));
-    view.rerender(<Host items={[]} sources={[importedSources[0]]} />);
-    expect(screen.getByTestId("canvas").dataset.selected ?? "").toBe("");
-    expect(screen.queryByText(importedReceipt.examples[0].excerpt)).toBeNull();
-    expect(screen.queryByText("查看已读内容")).toBeNull();
-    expect(sourceReceipt).toHaveBeenCalledTimes(1);
-  },
-);
+it("does not replace a newer memory choice when a source page settles", async () => {
+  const page = {
+    items: memories,
+    total: 2,
+    offset: 0,
+    limit: 20,
+    source_ids_by_memory: { m0: ["codex"], m1: ["codex"] },
+  };
+  const view = render(
+    <Host simplified sources={importedSources} libraryView={pagedView(page)} />,
+  );
+  await act(async () => {});
+  await userEvent.click(screen.getByRole("button", { name: "图中选择 Codex" }));
+  await userEvent.click(
+    screen.getByRole("button", { name: "图中选择 第二条" }),
+  );
+  expect(await screen.findByRole("heading", { name: "第二条" })).toBeTruthy();
 
-it("REV09: compact preview close keeps source selection for reopening", async () => {
+  view.rerender(
+    <Host
+      simplified
+      sources={importedSources}
+      libraryView={{
+        ...pagedView(page),
+        source: "codex",
+        settledSource: "codex",
+      }}
+    />,
+  );
+  expect(screen.getByRole("heading", { name: "第二条" })).toBeTruthy();
+});
+
+it("closing the compact source inspector cancels its pending auto-open", async () => {
   vi.stubGlobal("matchMedia", (query: string) => ({
     matches: query.includes("max-width"),
     addEventListener() {},
     removeEventListener() {},
   }));
-  sourceReceipt.mockResolvedValue(importedReceipt);
-  render(<Host items={[]} sources={importedSources} />);
-  await openSourceReceipt();
-  await screen.findByText(importedReceipt.examples[0].excerpt);
+  const page = {
+    items: memories,
+    total: 2,
+    offset: 0,
+    limit: 20,
+    source_ids_by_memory: { m0: ["codex"], m1: ["codex"] },
+  };
+  const view = render(
+    <Host simplified sources={importedSources} libraryView={pagedView(page)} />,
+  );
+  await act(async () => {});
+  await userEvent.click(screen.getByRole("button", { name: "图中选择 Codex" }));
   await userEvent.click(screen.getByRole("button", { name: "关闭记忆详情" }));
-  expect(document.querySelector(".source-import-receipt")).toBeNull();
+
+  view.rerender(
+    <Host
+      simplified
+      sources={importedSources}
+      libraryView={{
+        ...pagedView(page),
+        source: "codex",
+        settledSource: "codex",
+      }}
+    />,
+  );
+  expect(screen.queryByRole("dialog", { name: "记忆与原文" })).toBeNull();
+  expect(screen.queryByRole("heading", { name: "第一条" })).toBeNull();
+});
+
+it("keeps the source selection and offers retry when its page fails", async () => {
+  const onRetry = vi.fn();
+  const page = {
+    items: memories,
+    total: 2,
+    offset: 0,
+    limit: 20,
+    source_ids_by_memory: { m0: ["codex"], m1: ["codex"] },
+  };
+  const view = render(
+    <Host simplified sources={importedSources} libraryView={pagedView(page)} />,
+  );
+  await act(async () => {});
+  await userEvent.click(screen.getByRole("button", { name: "图中选择 Codex" }));
+  view.rerender(
+    <Host
+      simplified
+      sources={importedSources}
+      libraryView={{
+        ...pagedView(page),
+        source: "codex",
+        settledSource: "codex",
+        failed: true,
+        onRetry,
+      }}
+    />,
+  );
+  expect(
+    screen
+      .getAllByRole("alert")
+      .some((node) => node.textContent === "记忆暂时无法读取"),
+  ).toBe(true);
+  await userEvent.click(screen.getByRole("button", { name: "重试" }));
+  expect(onRetry).toHaveBeenCalledOnce();
+});
+
+it("drops a pending source selection when that source is revoked", async () => {
+  const page = {
+    items: memories,
+    total: 2,
+    offset: 0,
+    limit: 20,
+    source_ids_by_memory: { m0: ["codex"], m1: ["codex"] },
+  };
+  const view = render(
+    <Host
+      simplified
+      sources={[importedSources[0]]}
+      libraryView={pagedView(page)}
+    />,
+  );
+  await act(async () => {});
+  await userEvent.click(screen.getByRole("button", { name: "图中选择 Codex" }));
   expect(screen.getByTestId("canvas").dataset.selected).toBe("source:codex");
-  await userEvent.click(
-    screen.getByRole("button", { name: "查看选中来源原文" }),
+
+  view.rerender(
+    <Host
+      simplified
+      sources={[{ ...importedSources[0], authorized: false }]}
+      libraryView={{
+        ...pagedView(page),
+        source: "codex",
+        settledSource: "codex",
+      }}
+    />,
+  );
+  expect(screen.queryByRole("button", { name: "图中选择 Codex" })).toBeNull();
+  expect(screen.queryByTestId("canvas")).toBeNull();
+  expect(
+    screen.getByRole("list", { name: "地图记忆列表" }).children,
+  ).toHaveLength(0);
+  expect(screen.queryByText("正在读取记忆详情…")).toBeNull();
+
+  view.rerender(
+    <Host
+      simplified
+      sources={[importedSources[0]]}
+      libraryView={{
+        ...pagedView(page),
+        source: "codex",
+        settledSource: "codex",
+      }}
+    />,
+  );
+  expect(screen.getByTestId("canvas").dataset.selected).not.toBe(
+    "source:codex",
   );
-  expect(screen.getByRole("dialog", { name: "记忆与原文" })).toBeTruthy();
-  expect(screen.getByText("查看已读内容")).toBeTruthy();
 });
 
 it("REV09: first governed memories replace fallback without remounting the graph or breaking memory IDs", async () => {
@@ -1325,5 +1460,4 @@ it.each([
   render(<Host items={[]} sources={sources} />);
   expect(screen.queryByTestId("canvas")).toBeNull();
   expect(document.querySelector(".source-import-receipt")).toBeNull();
-  expect(sourceReceipt).not.toHaveBeenCalled();
 });
diff --git a/memhub/apps/web/src/MemoryMap.tsx b/memhub/apps/web/src/MemoryMap.tsx
index 4f897cd..cf5ae56 100644
--- a/memhub/apps/web/src/MemoryMap.tsx
+++ b/memhub/apps/web/src/MemoryMap.tsx
@@ -42,13 +42,13 @@ import { memoryText, useMemoryLanguage } from "./memory-language";
 import {
   type MapAxis,
   mapGroups,
+  omitRevokedSourceGraph,
   projectMap,
   projectPagedSourceMap,
 } from "./memory-map-model";
 import { OptionMenu } from "./OptionMenu";
 import { R6Button } from "./R6Controls";
 import { SelectedMemoryDetail } from "./SelectedMemoryDetail";
-import { SourceImportReceipt } from "./SourceImportReceipt";
 import { sourceNameOr } from "./source-label";
 import { useMediaPreference } from "./use-media-preference";
 import { Answer, NodeDetail, type SavedAnswer } from "./Workbench";
@@ -98,6 +98,8 @@ type Props = {
     failed: boolean;
     graphFailed?: boolean;
     stale: boolean;
+    settledSource?: string;
+    settledQuery?: string;
     /** The source id the SDK uses for "no source given" (shown as 未识别来源). */
     unattributedSource?: string;
     onQuery(value: string): void;
@@ -140,7 +142,7 @@ function monthDay(value: string, locale: GuiLocale): string | null {
 
 /** D96: when the cited conversation happened, not when the memory was generated. */
 function sourceStamp(
-  time: { at: string; basis: "conversation" | "import" } | undefined,
+  time: { at: string; basis: "conversation" | "import" | "write" } | undefined,
   fallback: string | null | undefined,
   locale: GuiLocale = "zh-CN",
 ): string {
@@ -148,7 +150,11 @@ function sourceStamp(
   if (time && day)
     return memoryText(
       locale,
-      time.basis === "conversation" ? "对话于 {day}" : "导入于 {day}",
+      time.basis === "conversation"
+        ? "对话于 {day}"
+        : time.basis === "write"
+          ? "写入于 {day}"
+          : "导入于 {day}",
       { day },
     );
   return dayStamp(fallback, locale);
@@ -179,20 +185,24 @@ export function MemoryMap(props: Props) {
   const library = props.libraryView;
   const serverPaged = library !== undefined;
   const unattributedLabel = text("未识别来源");
+  const authorizedBaseGraph = useMemo(
+    () => omitRevokedSourceGraph(baseGraph, sources),
+    [baseGraph, sources],
+  );
   const graph = useMemo(
     () =>
       serverPaged
         ? projectPagedSourceMap(
-            baseGraph,
+            authorizedBaseGraph,
             library?.page ?? null,
             sources,
             library?.unattributedSource
               ? { [library.unattributedSource]: unattributedLabel }
               : undefined,
           )
-        : baseGraph,
+        : authorizedBaseGraph,
     [
-      baseGraph,
+      authorizedBaseGraph,
       serverPaged,
       library?.page,
       library?.unattributedSource,
@@ -214,6 +224,9 @@ export function MemoryMap(props: Props) {
   }, [submittedQuery]);
   const [selectedId, setSelectedId] = useState<string | null>(null);
   const [groupNode, setGroupNode] = useState<MemoryGraphNode | null>(null);
+  const [pendingSourceSelection, setPendingSourceSelection] = useState<
+    string | null
+  >(null);
   const [question, setQuestion] = useState("");
   const [inspectorOpen, setInspectorOpen] = useState(false);
   const [filterNotice, setFilterNotice] = useState("");
@@ -231,10 +244,18 @@ export function MemoryMap(props: Props) {
       ? `source:${library.source}`
       : ""
     : groupKey;
-  const active = useMemo(
-    () => memories.filter((m) => m.status !== "deleted"),
-    [memories],
-  );
+  const active = useMemo(() => {
+    const revoked = new Set(
+      sources
+        .filter((source) => source.authorized === false)
+        .map((source) => source.source_id),
+    );
+    return memories.filter((memory) => {
+      if (memory.status === "deleted") return false;
+      const sourceIds = library?.page?.source_ids_by_memory?.[memory.memory_id];
+      return !sourceIds?.some((sourceId) => revoked.has(sourceId));
+    });
+  }, [library?.page?.source_ids_by_memory, memories, sources]);
   const arrival = resolveMemoryArrival({
     arrival: props.arrival,
     processing: props.processing,
@@ -323,7 +344,7 @@ export function MemoryMap(props: Props) {
     // to dominate that page. Keyword search is intentionally page-scoped.
     if (serverPaged && !library?.query) {
       return (
-        baseGraph ??
+        authorizedBaseGraph ??
         projectMap(
           graph,
           active,
@@ -364,7 +385,7 @@ export function MemoryMap(props: Props) {
         })),
     };
   }, [
-    baseGraph,
+    authorizedBaseGraph,
     graph,
     axis,
     filtered,
@@ -401,6 +422,7 @@ export function MemoryMap(props: Props) {
     // Source revocation also clears a suspended preview, without closing shared memory detail.
     if (groupNode && !visibleGroupNode) {
       setGroupNode(null);
+      setPendingSourceSelection(null);
       setInspectorOpen(false);
     }
   }, [groupNode, visibleGroupNode]);
@@ -427,6 +449,7 @@ export function MemoryMap(props: Props) {
   const pick = useCallback(
     (memoryId: string, scroll = true) => {
       if (!active.some((memory) => memory.memory_id === memoryId)) return;
+      setPendingSourceSelection(null);
       scrollSelection.current = scroll;
       if (!filtered.some((memory) => memory.memory_id === memoryId)) {
         setGroupKey("");
@@ -440,6 +463,32 @@ export function MemoryMap(props: Props) {
     },
     [active, filtered, onOpenMemory],
   );
+  useEffect(() => {
+    if (
+      !library ||
+      !pendingSourceSelection ||
+      overlaysSuspended ||
+      library.loading ||
+      library.failed ||
+      !selectedSource ||
+      selectedSource.source_id !== pendingSourceSelection ||
+      library.source !== pendingSourceSelection ||
+      (library.settledSource ?? library.source) !== pendingSourceSelection ||
+      (library.settledQuery ?? library.query).trim() ||
+      library.page?.offset !== 0
+    )
+      return;
+    setPendingSourceSelection(null);
+    const first = active[0];
+    if (first) pick(first.memory_id, false);
+  }, [
+    active,
+    library,
+    overlaysSuspended,
+    pendingSourceSelection,
+    pick,
+    selectedSource,
+  ]);
   useEffect(() => {
     if (!props.simplified || openedDefault.current || overlaysSuspended) return;
     if (library?.loading && filtered.length === 0) return;
@@ -490,6 +539,7 @@ export function MemoryMap(props: Props) {
   const clear = useCallback(() => {
     setSelectedId(null);
     setGroupNode(null);
+    setPendingSourceSelection(null);
     setInspectorOpen(false);
     onCloseDetail();
   }, [onCloseDetail]);
@@ -521,6 +571,7 @@ export function MemoryMap(props: Props) {
       clear();
       return;
     }
+    setPendingSourceSelection(null);
     setInspectorOpen(false);
   }
   function chooseGroup(key: string) {
@@ -547,8 +598,17 @@ export function MemoryMap(props: Props) {
       setAxis("source");
       setGroupKey(`source:${node.source_id}`);
       if (library) {
+        setPendingSourceSelection(node.source_id);
+        library.onQuery("");
         library.onSource(node.source_id);
         library.resetPage();
+      } else {
+        const first = sourceGroups
+          .find((entry) => entry.key === `source:${node.source_id}`)
+          ?.memoryIds.find((memoryId) =>
+            active.some((memory) => memory.memory_id === memoryId),
+          );
+        if (first) pick(first, false);
       }
     }
     if (node.node_type === "topic") {
@@ -1025,11 +1085,13 @@ export function MemoryMap(props: Props) {
             <button
               type="button"
               className="quiet map-tool"
-              aria-label={text(
-                selectedSource ? "查看选中来源原文" : "查看选中记忆原文",
-              )}
+              aria-label={text("记忆详情")}
               title={text("查看原文")}
-              onClick={() => setInspectorOpen(true)}
+              onClick={() => {
+                setInspectorOpen(true);
+                if (selectedSource && library)
+                  setPendingSourceSelection(selectedSource.source_id);
+              }}
             >
               <FileText size={17} />
             </button>
@@ -1173,7 +1235,7 @@ export function MemoryMap(props: Props) {
             {inspectorDetail
               ? text("记忆详情")
               : selectedSource
-                ? text("来源原文")
+                ? text("记忆详情")
                 : visibleGroupNode
                   ? text("关联节点")
                   : text("记忆详情")}
@@ -1192,13 +1254,33 @@ export function MemoryMap(props: Props) {
         </header>
         <div className="map-inspector-body" ref={inspectorBody}>
           {selectedSource ? (
-            inspectorOpen &&
-            !overlaysSuspended && (
-              <SourceImportReceipt
-                key={selectedSource.source_id}
-                source={selectedSource}
-              />
-            )
+            <div className="map-inspector-empty map-inspector-loading">
+              {library?.failed &&
+              library.source === selectedSource.source_id ? (
+                <>
+                  <p role="alert">{text("记忆暂时无法读取")}</p>
+                  <button
+                    type="button"
+                    className="secondary"
+                    onClick={library.onRetry}
+                  >
+                    {text("重试")}
+                  </button>
+                </>
+              ) : pendingSourceSelection || library?.loading ? (
+                <>
+                  <div className="map-detail-skeleton" aria-hidden="true">
+                    <i />
+                    <i />
+                    <i />
+                    <i />
+                  </div>
+                  <p role="status">{text("正在读取记忆详情…")}</p>
+                </>
+              ) : (
+                <p>{text("还没有已生成的记忆")}</p>
+              )}
+            </div>
           ) : props.simplified && inspectorDetail ? (
             <div
               aria-busy={
diff --git a/memhub/apps/web/src/MemorySpace.tsx b/memhub/apps/web/src/MemorySpace.tsx
index 9381265..5c57097 100644
--- a/memhub/apps/web/src/MemorySpace.tsx
+++ b/memhub/apps/web/src/MemorySpace.tsx
@@ -1046,6 +1046,8 @@ export function MemorySpace({
                 failed: sdkLibrary.failed,
                 graphFailed: sdkGraphFailed,
                 stale: !sdkGraphSettled,
+                settledSource: sdkLibrary.settledSource,
+                settledQuery: sdkLibrary.settledQuery,
                 unattributedSource: UNATTRIBUTED_SOURCE_ID,
                 onQuery: setSdkLibraryQuery,
                 onSource: setSdkLibrarySource,
diff --git a/memhub/apps/web/src/SelectedMemoryDetail.test.tsx b/memhub/apps/web/src/SelectedMemoryDetail.test.tsx
index 3964c72..4409ff5 100644
--- a/memhub/apps/web/src/SelectedMemoryDetail.test.tsx
+++ b/memhub/apps/web/src/SelectedMemoryDetail.test.tsx
@@ -176,3 +176,41 @@ it("stamps today's source and conversation time without an evidence page", () =>
   expect(screen.getByText(/对话时间 今天 /)).toBeTruthy();
   expect(screen.queryByRole("button", { name: "查看原文" })).toBeNull();
 });
+it("shows an experience's conversation time when the SDK stored one", () => {
+  const today = new Date().toISOString();
+  render(
+    <SelectedMemoryDetail
+      detail={{
+        ...detail,
+        evidence: [],
+        sdk_status: "active",
+        source_id: "codebuddy",
+        sdk_memory_at: today,
+        sdk_created_at: today,
+        sdk_ref: { scene: "coding", kind: "experience", variant: "coding" },
+      }}
+      sources={[]}
+    />,
+  );
+  expect(screen.getByText(/对话时间 今天 /)).toBeTruthy();
+  expect(screen.queryByText(/时间未知/)).toBeNull();
+});
+it("labels an experience without a conversation time by its write time", () => {
+  render(
+    <SelectedMemoryDetail
+      detail={{
+        ...detail,
+        evidence: [],
+        sdk_status: "active",
+        source_id: "codebuddy",
+        sdk_memory_at: null,
+        sdk_created_at: new Date().toISOString(),
+        sdk_ref: { scene: "working", kind: "experience", variant: "working" },
+      }}
+      sources={[]}
+    />,
+  );
+  expect(screen.getByText(/来源 codebuddy · 写入时间 今天 /)).toBeTruthy();
+  expect(screen.queryByText(/对话时间/)).toBeNull();
+  expect(screen.queryByText(/时间未知/)).toBeNull();
+});
diff --git a/memhub/apps/web/src/SelectedMemoryDetail.tsx b/memhub/apps/web/src/SelectedMemoryDetail.tsx
index 6b32b32..10d2ace 100644
--- a/memhub/apps/web/src/SelectedMemoryDetail.tsx
+++ b/memhub/apps/web/src/SelectedMemoryDetail.tsx
@@ -19,7 +19,10 @@ export function SelectedMemoryDetail({
   detail: MemoryDetail;
   sources: SourceResponse[];
   /** D99: when the cited conversation happened, in the list's own terms. */
-  conversationAt?: { at: string; basis: "conversation" | "import" } | null;
+  conversationAt?: {
+    at: string;
+    basis: "conversation" | "import" | "write";
+  } | null;
 }) {
   const { locale, t } = useGuiLanguage();
   const memory = detail.memory;
@@ -42,6 +45,17 @@ export function SelectedMemoryDetail({
           ? t("selectedMemory.status.superseded")
           : t("selectedMemory.status.unknown");
   const sourceId = detail.source_id ?? evidence?.source_id;
+  const created = clockStamp(
+    detail.sdk_created_at ?? memory.updated_at,
+    locale,
+  );
+  // An SDK experience or fact stored before hy-memory 2.0.7 has no
+  // conversation time. It shows when it was written, named as such, instead
+  // of "时间未知".
+  const writtenOnly =
+    detail.sdk_ref !== undefined &&
+    detail.sdk_ref.kind !== "chat" &&
+    !detail.sdk_memory_at;
   const source =
     detail.sdk_status && sourceId === UNATTRIBUTED_SOURCE_ID
       ? t("selectedMemory.sourceUnattributed")
@@ -77,25 +91,28 @@ export function SelectedMemoryDetail({
         <p className="selected-memory-source">
           {t("selectedMemory.sourceLine", {
             source,
-            time: clockStamp(
-              detail.sdk_memory_at ?? evidence?.created_at,
-              locale,
-            ),
+            time: writtenOnly
+              ? t("selectedMemory.writeTime", { created })
+              : clockStamp(
+                  detail.sdk_memory_at ?? evidence?.created_at,
+                  locale,
+                ),
           })}
         </p>
       ) : null}
-      <p className="selected-memory-times">
-        {t("selectedMemory.timestamps", {
-          conversation: clockStamp(
-            conversationAt?.at ?? detail.sdk_memory_at ?? evidence?.created_at,
-            locale,
-          ),
-          created: clockStamp(
-            detail.sdk_created_at ?? memory.updated_at,
-            locale,
-          ),
-        })}
-      </p>
+      {writtenOnly ? null : (
+        <p className="selected-memory-times">
+          {t("selectedMemory.timestamps", {
+            conversation: clockStamp(
+              conversationAt?.at ??
+                detail.sdk_memory_at ??
+                evidence?.created_at,
+              locale,
+            ),
+            created,
+          })}
+        </p>
+      )}
     </article>
   );
 }
diff --git a/memhub/apps/web/src/debug-log-types.ts b/memhub/apps/web/src/debug-log-types.ts
index 0df67b0..f3da192 100644
--- a/memhub/apps/web/src/debug-log-types.ts
+++ b/memhub/apps/web/src/debug-log-types.ts
@@ -4,14 +4,7 @@ export interface DebugLogEntry {
   id: string;
   timestamp: string;
   level: DebugLogLevel;
-  service:
-    | "runtime"
-    | "core"
-    | "host-bridge"
-    | "desktop"
-    | "cli"
-    | "hy-memory"
-    | "renderer";
+  service: "runtime" | "core" | "host-bridge" | "desktop" | "cli" | "renderer";
   component: string;
   event: string;
   outcome: string;
@@ -21,7 +14,7 @@ export interface DebugLogEntry {
   trace_id?: string;
   details?: Record<string, string | number | boolean | null>;
   message?: string;
-  source: "diagnostic" | "stdout" | "stderr" | "console" | "pipeline";
+  source: "diagnostic" | "stdout" | "stderr" | "console";
 }
 
 export interface DebugLogSnapshot {
diff --git a/memhub/apps/web/src/memory-map-model.test.ts b/memhub/apps/web/src/memory-map-model.test.ts
index 5889d4e..7fd7991 100644
--- a/memhub/apps/web/src/memory-map-model.test.ts
+++ b/memhub/apps/web/src/memory-map-model.test.ts
@@ -7,6 +7,7 @@ import type {
 import { describe, expect, it } from "vitest";
 import {
   mapGroups,
+  omitRevokedSourceGraph,
   projectMap,
   projectPagedSourceMap,
 } from "./memory-map-model";
@@ -95,6 +96,18 @@ describe("memory map membership", () => {
     expect(result.nodes.map((n) => n.node_id)).not.toContain("memory:c");
     expect(result.edges).toHaveLength(2);
   });
+  it("removes revoked source subgraphs while retaining authorized ones", () => {
+    const result = omitRevokedSourceGraph(graph, [
+      { source_id: "codex", authorized: false },
+      { source_id: "cursor", authorized: true },
+    ] as SourceResponse[]);
+    expect(result?.nodes.map((node) => node.node_id)).toEqual([
+      "memory:c",
+      "source:cursor",
+    ]);
+    expect(result?.edges.map((edge) => edge.edge_id)).toEqual(["c"]);
+    expect(result?.projected_memories).toBe(1);
+  });
   it("builds topic edges only from actual tags, without creating new facts", () => {
     const result = projectMap(graph, memories, "topic", ["a", "b"]);
     expect(result.nodes.filter((n) => n.node_type === "memory")).toHaveLength(
diff --git a/memhub/apps/web/src/memory-map-model.ts b/memhub/apps/web/src/memory-map-model.ts
index fe14c26..a84308f 100644
--- a/memhub/apps/web/src/memory-map-model.ts
+++ b/memhub/apps/web/src/memory-map-model.ts
@@ -15,6 +15,85 @@ export type MapGroup = {
   sourceCount: number;
 };
 
+/** Remove every graph fact that depends on an explicitly revoked source.
+ * Unknown source IDs remain visible because they are not an authorization denial.
+ */
+export function omitRevokedSourceGraph(
+  graph: MemoryGraphResponse | null,
+  sources: SourceResponse[],
+): MemoryGraphResponse | null {
+  if (!graph) return null;
+  const revoked = new Set(
+    sources
+      .filter((source) => source.authorized === false)
+      .map((source) => source.source_id),
+  );
+  if (!revoked.size) return graph;
+
+  const nodes = new Map(graph.nodes.map((node) => [node.node_id, node]));
+  const blocked = new Set(
+    graph.nodes
+      .filter(
+        (node) =>
+          node.node_type === "source" &&
+          !!node.source_id &&
+          revoked.has(node.source_id),
+      )
+      .map((node) => node.node_id),
+  );
+  for (const edge of graph.edges) {
+    if (edge.origin !== "provenance") continue;
+    const ends = [edge.source_node_id, edge.target_node_id];
+    if (!ends.some((id) => blocked.has(id))) continue;
+    for (const id of ends) {
+      if (nodes.get(id)?.node_type === "memory") blocked.add(id);
+    }
+  }
+
+  const candidateEdges = graph.edges.filter(
+    (edge) =>
+      !blocked.has(edge.source_node_id) && !blocked.has(edge.target_node_id),
+  );
+  const retained = new Set(
+    graph.nodes
+      .filter(
+        (node) => node.node_type === "memory" && !blocked.has(node.node_id),
+      )
+      .map((node) => node.node_id),
+  );
+  let changed = true;
+  while (changed) {
+    changed = false;
+    for (const edge of candidateEdges) {
+      if (
+        retained.has(edge.source_node_id) &&
+        !retained.has(edge.target_node_id)
+      ) {
+        retained.add(edge.target_node_id);
+        changed = true;
+      } else if (
+        retained.has(edge.target_node_id) &&
+        !retained.has(edge.source_node_id)
+      ) {
+        retained.add(edge.source_node_id);
+        changed = true;
+      }
+    }
+  }
+  const retainedNodes = graph.nodes.filter((node) => retained.has(node.node_id));
+  return {
+    ...graph,
+    projected_memories: retainedNodes.filter(
+      (node) => node.node_type === "memory",
+    ).length,
+    nodes: retainedNodes,
+    edges: candidateEdges.filter(
+      (edge) =>
+        retained.has(edge.source_node_id) && retained.has(edge.target_node_id),
+    ),
+  };
+}
+
 /** The current authorized page owns provenance; the global graph may be truncated.
  * Callers must pass null while page authority is stale. Catalog metadata names
  * sources and can veto revoked ones, but never invents evidence membership.
diff --git a/memhub/apps/web/src/sdk-memory-map.test.ts b/memhub/apps/web/src/sdk-memory-map.test.ts
index 55d8667..0b18c92 100644
--- a/memhub/apps/web/src/sdk-memory-map.test.ts
+++ b/memhub/apps/web/src/sdk-memory-map.test.ts
@@ -119,8 +119,12 @@ it("carries each memory's scene and namespace, and edits an experience's solutio
   });
   expect(page.window_id).toBe("w1");
   expect(page.window_reset).toBe(true);
-  // An experience has no conversation time to show as "imported".
-  expect(page.source_time_by_memory?.e).toBeUndefined();
+  // Stored without a conversation time: it shows when it was written, not
+  // an "import".
+  expect(page.source_time_by_memory?.e).toEqual({
+    at: new Date((NOW - 86_400) * 1_000).toISOString(),
+    basis: "write",
+  });
   expect(page.items[1]).toMatchObject({
     kind: "working",
     updated_at: new Date((NOW - 3_600) * 1_000).toISOString(),
@@ -140,3 +144,29 @@ it("carries each memory's scene and namespace, and edits an experience's solutio
   expect(sdkEditableText(experience)).toBe("Pin seed");
   expect(sdkEditableText(record({}))).toBe("memory");
 });
+
+it("shows an experience's conversation time when the SDK stored one", () => {
+  const at = NOW - 2 * 86_400;
+  const experience = record({
+    memory_id: "e",
+    layer: "experience",
+    scene: "coding",
+    kind: "experience",
+    variant: "coding",
+    memory_at: at,
+    created_at: NOW - 3_600,
+  });
+  const page = sdkLibraryPage({
+    memories: [experience],
+    total: 1,
+    limit: 20,
+    offset: 0,
+  });
+  expect(page.source_time_by_memory?.e).toEqual({
+    at: new Date(at * 1_000).toISOString(),
+    basis: "conversation",
+  });
+  expect(sdkDetail(experience).sdk_memory_at).toBe(
+    new Date(at * 1_000).toISOString(),
+  );
+});
diff --git a/memhub/apps/web/src/sdk-memory-map.ts b/memhub/apps/web/src/sdk-memory-map.ts
index 8e48829..ba13781 100644
--- a/memhub/apps/web/src/sdk-memory-map.ts
+++ b/memhub/apps/web/src/sdk-memory-map.ts
@@ -154,22 +154,24 @@ export function sdkDetail(record: SdkMemoryRecord): MemoryDetail {
 export function sdkLibraryPage(page: SdkMemoryPage): MemoryLibraryPage {
   const sourceTime: Record<
     string,
-    { at: string; basis: "conversation" | "import" }
+    { at: string; basis: "conversation" | "import" | "write" }
   > = {};
   const sourceIds: Record<string, string[]> = {};
   const refs: Record<string, SdkMemoryRef> = {};
   for (const record of page.memories) {
     sourceIds[record.memory_id] = [sourceOf(record)];
     refs[record.memory_id] = sdkMemoryRef(record);
-    // An experience or fact has no conversation time; the list shows when it
-    // was last updated instead of calling its creation an import.
-    if (refs[record.memory_id].kind !== "chat") continue;
+    // An experience or fact stored before hy-memory 2.0.7 has no
+    // conversation time. It shows when it was written, labelled as such,
+    // rather than calling its creation an import.
+    const chat = refs[record.memory_id].kind === "chat";
     const conversation = plausible(record.memory_at);
-    const at = conversation ?? plausible(record.gmt_created);
+    const at =
+      conversation ?? plausible(chat ? record.gmt_created : record.created_at);
     if (at === null) continue;
     sourceTime[record.memory_id] = {
       at: new Date(at * 1_000).toISOString(),
-      basis: conversation !== null ? "conversation" : "import",
+      basis: conversation !== null ? "conversation" : chat ? "import" : "write",
     };
   }
   return {
diff --git a/memhub/docs/ARCHITECTURE.md b/memhub/docs/ARCHITECTURE.md
index c6af59e..26d05dc 100644
--- a/memhub/docs/ARCHITECTURE.md
+++ b/memhub/docs/ARCHITECTURE.md
@@ -282,7 +282,7 @@ Support is declared per immutable GitHub Release manifest:
 - Linux x64/arm64 Runtime on native GHA runners.
 - macOS arm64/x64 App/Runtime: Developer ID, notarization, native acceptance.
 - Windows is not supported.
-- `hy-memory==2.0.6` is the current stable SDK dependency; the release
+- `hy-memory==2.0.8` is the current stable SDK dependency; the release
   lock and [`../MEMORY.md`](../MEMORY.md) are authoritative when the pin changes.
 
 See [`INSTALLATION.md`](INSTALLATION.md).
diff --git a/memhub/docs/PRODUCT_LOG_TROUBLESHOOTING.md b/memhub/docs/PRODUCT_LOG_TROUBLESHOOTING.md
index 94af606..66cd356 100644
--- a/memhub/docs/PRODUCT_LOG_TROUBLESHOOTING.md
+++ b/memhub/docs/PRODUCT_LOG_TROUBLESHOOTING.md
@@ -7,12 +7,12 @@
 3. 如果停住，停在导入、模型调用还是最终写入。
 
 默认 `info` 下的官方结构化事件不应写入会话正文、模型输入输出或 API Key；但本地原始日志允许
-新增受限标量标签，仍应按私人数据处理，不能直接对外发送。`debug` 会额外启用 HY pipeline
-trace，其中可能包含完整 prompt/response 和获授权的对话正文。
+新增受限标量标签，仍应按私人数据处理，不能直接对外发送。`debug` 会让 HY 的
+`hy_memory.log` 记录 DEBUG 行，其中可能包含完整 prompt/response 和获授权的对话正文。
 
 排查时只通过 App 的“导出日志”提供文件：导出过程会按安全白名单重新生成诊断事件，不会打包
-原始日志。不要直接发送 `diagnostic-*.jsonl`、整个 `~/.one-memory` 数据目录或 debug pipeline
-日志；后者如确需分享，必须由开发人员另行确认并脱敏。
+原始日志。不要直接发送 `diagnostic-*.jsonl`、整个 `~/.one-memory` 数据目录或 debug 级别的
+`hy_memory.log`；后者如确需分享，必须由开发人员另行确认并脱敏。
 
 ## 1. Log level 设什么
 
@@ -181,8 +181,8 @@ memhub logs --since 24h --errors-only --tail 500 --json
 memhub status --json
 ```
 
-以下 `jq` 命令仅用于 `diagnostic-core.jsonl`，不要对 `core.log`、`service.log` 或 debug
-pipeline 日志直接运行后把完整输出发给他人。
+以下 `jq` 命令仅用于 `diagnostic-core.jsonl`，不要对 `core.log`、`service.log` 或
+`hy_memory.log` 直接运行后把完整输出发给他人。
 
 ### 查看最近的失败
 
diff --git a/memhub/docs/contracts/one-memory-recall-install-v1.md b/memhub/docs/contracts/one-memory-recall-install-v1.md
index 9fa6c8c..69d0c35 100644
--- a/memhub/docs/contracts/one-memory-recall-install-v1.md
+++ b/memhub/docs/contracts/one-memory-recall-install-v1.md
@@ -39,10 +39,12 @@ which talks to a broker on the same machine.
 | `ONE_MEMORY_BROKER_DISCOVERY` | absolute path to the discovery file | How the server finds the running app. Wrong value means recall returns nothing, with no error. |
 | `ONE_MEMORY_MANAGED_BY` | `one-memory` | Marks the entry as ours, so our installer will update or remove it later instead of treating it as a stranger's and leaving it alone. |
 | `ELECTRON_RUN_AS_NODE` | `1` | The MCP entry ships inside the app bundle and is started through the app's own binary. Without this the binary opens the window instead of running the script. |
+| `ONE_MEMORY_MCP_TOOLSET` | tool set version, e.g. `3` | Names the MemHub tool set the entry was written for. MemHub changes it only when tools are added, removed or renamed, so hosts that cache a server's tool list by its launch config restart the server and drop removed tools. The server itself ignores it. |
 
 `ONE_MEMORY_MANAGED_BY` is optional for a hand-made entry. Set it if you want
 MemHub to manage the entry for you afterwards; leave it out if you would
-rather own the entry yourself.
+rather own the entry yourself. `ONE_MEMORY_MCP_TOOLSET` is optional too; a
+managed entry without the current value is rewritten on the next upgrade.
 
 ## 3. MCP hosts
 
@@ -57,7 +59,8 @@ Register a stdio server named `one-memory`. The product name is MemHub; the conf
       "env": {
         "ELECTRON_RUN_AS_NODE": "1",
         "ONE_MEMORY_BROKER_DISCOVERY": "<absolute path to the discovery file>",
-        "ONE_MEMORY_MANAGED_BY": "one-memory"
+        "ONE_MEMORY_MANAGED_BY": "one-memory",
+        "ONE_MEMORY_MCP_TOOLSET": "3"
       }
     }
   }
diff --git a/memhub/package.json b/memhub/package.json
index 2bd53b6..954a77c 100644
--- a/memhub/package.json
+++ b/memhub/package.json
@@ -2,7 +2,7 @@
   "name": "one-memory",
   "version": "1.0.5",
   "release": {
-    "build_number": 33,
+    "build_number": 35,
     "macos_min_version": "13.0",
     "app_update": "manual",
     "runtime_manifest_schema": 1,
diff --git a/memhub/packages/contracts/src/index.ts b/memhub/packages/contracts/src/index.ts
index 2ca034a..1d443ba 100644
--- a/memhub/packages/contracts/src/index.ts
+++ b/memhub/packages/contracts/src/index.ts
@@ -468,10 +468,12 @@ export interface MemoryLibraryPage {
   limit: number;
   /** Active evidence sources for each returned memory, from the same authorized page snapshot. */
   source_ids_by_memory?: Record<string, string[]>;
-  /** When each returned memory's cited conversation happened (D96). `import` when only the import time is known. */
+  /** When each returned memory's cited conversation happened (D96). `import`
+   * when only the import time is known; `write` for an SDK experience or fact
+   * stored without a conversation time, which shows when it was written. */
   source_time_by_memory?: Record<
     string,
-    { at: string; basis: "conversation" | "import" }
+    { at: string; basis: "conversation" | "import" | "write" }
   >;
   /** True when the read budget ran out mid-page (typically during a history
    * import): a partial snapshot, not an authoritative view. Show "importing"
diff --git a/memhub/packages/host-bridge/src/install-contract.test.ts b/memhub/packages/host-bridge/src/install-contract.test.ts
index 7d67eae..f67e778 100644
--- a/memhub/packages/host-bridge/src/install-contract.test.ts
+++ b/memhub/packages/host-bridge/src/install-contract.test.ts
@@ -5,6 +5,8 @@ import { describe, expect, it } from "vitest";
 import type { RecallRuntime } from "./recall-installer.js";
 import {
   DISCOVERY_KEY,
+  MCP_TOOLSET_KEY,
+  MCP_TOOLSET_VERSION,
   OWNER_KEY,
   OWNER_VALUE,
   renderConnectionMaterial,
@@ -60,6 +62,7 @@ describe("mcp install snippet", () => {
     const env = parsed.mcpServers[SERVER_NAME].env;
     expect(env[DISCOVERY_KEY]).toBe(runtime().discoveryPath);
     expect(env[OWNER_KEY]).toBe(OWNER_VALUE);
+    expect(env[MCP_TOOLSET_KEY]).toBe(MCP_TOOLSET_VERSION);
     expect(env.ELECTRON_RUN_AS_NODE).toBe("1");
   });
 
@@ -174,9 +177,10 @@ describe("public contract document", () => {
     // reports it to us. Same class of problem R-33 treated for skill text,
     // and treated the same way: one source, checked by a test.
     const doc = await fs.readFile(CONTRACT_DOC, "utf8");
-    for (const key of [DISCOVERY_KEY, OWNER_KEY]) {
+    for (const key of [DISCOVERY_KEY, OWNER_KEY, MCP_TOOLSET_KEY]) {
       expect(doc).toContain(key);
     }
+    expect(doc).toContain(`"${MCP_TOOLSET_KEY}": "${MCP_TOOLSET_VERSION}"`);
     expect(doc).toContain(OWNER_VALUE);
     expect(doc).toContain(SERVER_NAME);
   });
diff --git a/memhub/packages/host-bridge/src/mcp-granted-tools.test.ts b/memhub/packages/host-bridge/src/mcp-granted-tools.test.ts
index 3ac0305..cc625a2 100644
--- a/memhub/packages/host-bridge/src/mcp-granted-tools.test.ts
+++ b/memhub/packages/host-bridge/src/mcp-granted-tools.test.ts
@@ -1,6 +1,7 @@
 import { describe, expect, it } from "vitest";
 
 import { type GrantedTools, McpSession } from "./mcp-session.js";
+import { MCP_TOOL_NAMES, MCP_TOOLSET_VERSION } from "./mcp-toolset.js";
 
 function sessionWith(granted: GrantedTools): McpSession {
   return new McpSession(
@@ -71,3 +72,14 @@ describe("BUG-15: the listed tools follow what the grant allows", () => {
     expect(names).toHaveLength(3);
   });
 });
+
+describe("the tool set stamped on installed entries", () => {
+  it("names exactly the tools the server lists", async () => {
+    // Change a tool without moving MCP_TOOLSET_VERSION and hosts that cache
+    // tool lists by config fingerprint keep showing the old set. Update both,
+    // and this pinned version, together.
+    const names = await listedNames(sessionWith(async () => null));
+    expect(names).toEqual([...MCP_TOOL_NAMES]);
+    expect(MCP_TOOLSET_VERSION).toBe("3");
+  });
+});
diff --git a/memhub/packages/host-bridge/src/mcp-session.ts b/memhub/packages/host-bridge/src/mcp-session.ts
index 5626593..2c81095 100644
--- a/memhub/packages/host-bridge/src/mcp-session.ts
+++ b/memhub/packages/host-bridge/src/mcp-session.ts
@@ -127,6 +127,8 @@ export class McpSession {
   }
 
   private async listedTools() {
+    // Any change to this set moves MCP_TOOL_NAMES and MCP_TOOLSET_VERSION in
+    // `mcp-toolset.ts`, so installed entries change and hosts relist.
     const definitions = [
       recallToolDefinition(),
       writeToolDefinition(),
diff --git a/memhub/packages/host-bridge/src/mcp-toolset.ts b/memhub/packages/host-bridge/src/mcp-toolset.ts
new file mode 100644
index 0000000..050ab9c
--- /dev/null
+++ b/memhub/packages/host-bridge/src/mcp-toolset.ts
@@ -0,0 +1,29 @@
+/** The tool contract of MemHub's MCP server, and the stamp that names it.
+ *
+ * Kept apart from `mcp-session.ts` so the installer can stamp an entry without
+ * pulling in the broker client, and listed here rather than inferred so a test
+ * can hold `tools/list` to it: change a tool there and that test fails until
+ * this list, and with it the version, moves too.
+ */
+
+/** The tools `tools/list` returns, in order. */
+export const MCP_TOOL_NAMES = [
+  "MemHub_recall",
+  "MemHub_write",
+  "MemHub_diagnostic_trace",
+] as const;
+
+/** Bumped whenever a tool is added, removed or renamed.
+ *
+ * Written into every managed MCP entry under `MCP_TOOLSET_KEY`. Hosts such as
+ * WorkBuddy cache a server's tool list under a fingerprint of its launch
+ * config, and our entry is otherwise byte-identical across upgrades, since it
+ * runs through the `current` symlink. Moving this value is what changes that
+ * fingerprint, so those hosts restart the server and drop the tools it no
+ * longer has. Deliberately not the build number: that would rewrite every
+ * host's config on every upgrade, for nothing.
+ */
+export const MCP_TOOLSET_VERSION = "3";
+
+/** The env key the version is written under. */
+export const MCP_TOOLSET_KEY = "ONE_MEMORY_MCP_TOOLSET";
diff --git a/memhub/packages/host-bridge/src/opencode-installer.test.ts b/memhub/packages/host-bridge/src/opencode-installer.test.ts
index 5d83d9a..931a37a 100644
--- a/memhub/packages/host-bridge/src/opencode-installer.test.ts
+++ b/memhub/packages/host-bridge/src/opencode-installer.test.ts
@@ -5,6 +5,7 @@ import { describe, expect, it } from "vitest";
 
 import {
   applyRecallInstall,
+  MCP_TOOLSET_VERSION,
   mcpConfigPath,
   opencodeLayout,
   planRecallInstall,
@@ -46,6 +47,7 @@ describe("OpenCode recall installer", () => {
         ELECTRON_RUN_AS_NODE: "1",
         ONE_MEMORY_MANAGED_BY: "one-memory",
         ONE_MEMORY_BROKER_DISCOVERY: "/opt/one-memory/runtime.json",
+        ONE_MEMORY_MCP_TOOLSET: MCP_TOOLSET_VERSION,
       },
       enabled: true,
     });
diff --git a/memhub/packages/host-bridge/src/recall-installer.ts b/memhub/packages/host-bridge/src/recall-installer.ts
index e4627ec..1f0cd23 100644
--- a/memhub/packages/host-bridge/src/recall-installer.ts
+++ b/memhub/packages/host-bridge/src/recall-installer.ts
@@ -58,6 +58,7 @@ import {
   readDshServer,
   removeDshServer,
 } from "./dsh-patch.js";
+import { MCP_TOOLSET_KEY, MCP_TOOLSET_VERSION } from "./mcp-toolset.js";
 import {
   RECALL_PRESENTATION_GUIDANCE,
   RECALL_PRESENTATION_GUIDANCE_ZH,
@@ -88,8 +89,12 @@ export {
   type RecallLinkState,
   type RecallProbe,
 } from "./recall-probe.js";
-
-export { planRecallRepoint, type RecallRepoint };
+export {
+  MCP_TOOLSET_KEY,
+  MCP_TOOLSET_VERSION,
+  planRecallRepoint,
+  type RecallRepoint,
+};
 
 /** Rewrite hosts whose installed command is not the one we would write now.
  *
@@ -2844,6 +2849,9 @@ function managedServer(runtime: RecallRuntime): ManagedServer {
       ELECTRON_RUN_AS_NODE: "1",
       [DISCOVERY_KEY]: runtime.discoveryPath,
       [OWNER_KEY]: OWNER_VALUE,
+      // Moves the host's fingerprint of this entry when the tool set changes;
+      // see `MCP_TOOLSET_VERSION`. An entry without it reads as stale.
+      [MCP_TOOLSET_KEY]: MCP_TOOLSET_VERSION,
       ...(runtime.credentialFile === undefined
         ? {}
         : { ONE_MEMORY_HOST_CREDENTIAL_FILE: runtime.credentialFile }),
diff --git a/memhub/packages/host-bridge/src/skill-drift.test.ts b/memhub/packages/host-bridge/src/skill-drift.test.ts
index 7d18fb3..cd681a1 100644
--- a/memhub/packages/host-bridge/src/skill-drift.test.ts
+++ b/memhub/packages/host-bridge/src/skill-drift.test.ts
@@ -5,6 +5,7 @@ import path from "node:path";
 import { describe, expect, it } from "vitest";
 
 import { applyCursorInstall, planCursorInstall } from "./cursor-installer.js";
+import { MCP_TOOLSET_KEY, MCP_TOOLSET_VERSION } from "./mcp-toolset.js";
 import {
   applyRecallInstall,
   applyRecallUninstall,
@@ -15,6 +16,7 @@ import {
   planRecallUninstall,
   SKILL_MARKER,
   SKILL_VERSION,
+  verifyRecallInstall,
 } from "./recall-installer.js";
 import {
   checkSkillDrift,
@@ -387,6 +389,59 @@ describe("skill drift (R-33)", () => {
     });
   }, 30_000);
 
+  it("marks stale an MCP entry written before the tool set was stamped, and sync stamps it", async () => {
+    // Our entry is otherwise identical across upgrades, so hosts that cache a
+    // tool list by config fingerprint keep the removed tools until it changes.
+    const cursorDir = await fixtureDir();
+    const discovery = path.join(cursorDir, "broker.json");
+    await writeFile(discovery, "{}\n");
+    await applyCursorInstall(
+      await planCursorInstall(options(cursorDir, discovery)),
+    );
+    const layout = cursorLayout(cursorDir);
+    const configPath = layout.configPath as string;
+    const config = JSON.parse(await readFile(configPath, "utf8")) as {
+      mcpServers: Record<string, { env: Record<string, string> }>;
+    };
+    expect(config.mcpServers["one-memory"].env[MCP_TOOLSET_KEY]).toBe(
+      MCP_TOOLSET_VERSION,
+    );
+    // What an earlier release wrote and journaled: the same entry, no stamp.
+    delete config.mcpServers["one-memory"].env[MCP_TOOLSET_KEY];
+    await writeFile(configPath, `${JSON.stringify(config, null, 2)}\n`);
+    const journal = JSON.parse(await readFile(layout.journalPath, "utf8"));
+    journal.installed_server_hash = createHash("sha256")
+      .update(JSON.stringify(config.mcpServers["one-memory"]))
+      .digest("hex");
+    await writeFile(layout.journalPath, `${JSON.stringify(journal)}\n`);
+
+    // Still intact as installed, so the link is not reported broken...
+    await expect(verifyRecallInstall(layout)).resolves.toMatchObject({
+      hostId: "cursor",
+    });
+    // ...but stale against this release, and safe to sync without asking.
+    await expect(
+      checkSkillDrift(layout, runtime(discovery)),
+    ).resolves.toMatchObject({
+      state: "stale",
+      detail: "the MCP entry predates the tool set this MemHub serves",
+    });
+    const sync = await planSkillSync(layout, runtime(discovery));
+    expect(sync?.requiresConfirmation).toBe(false);
+    if (sync === null) throw new Error("expected a stale entry sync");
+    await applyRecallInstall(sync.plan);
+
+    const synced = JSON.parse(await readFile(configPath, "utf8")) as {
+      mcpServers: Record<string, { env: Record<string, string> }>;
+    };
+    expect(synced.mcpServers["one-memory"].env[MCP_TOOLSET_KEY]).toBe(
+      MCP_TOOLSET_VERSION,
+    );
+    await expect(
+      checkSkillDrift(layout, runtime(discovery)),
+    ).resolves.toMatchObject({ state: "in_sync" });
+  }, 30_000);
+
   it("requires recall for dependent answers, not on every unrelated turn", async () => {
     const cursorDir = await fixtureDir();
     const discovery = path.join(cursorDir, "broker.json");
diff --git a/memhub/packages/host-bridge/src/skill-drift.ts b/memhub/packages/host-bridge/src/skill-drift.ts
index 9ea7758..9501d6c 100644
--- a/memhub/packages/host-bridge/src/skill-drift.ts
+++ b/memhub/packages/host-bridge/src/skill-drift.ts
@@ -2,6 +2,7 @@ import { createHash } from "node:crypto";
 import path from "node:path";
 import { captureFilePath, namesCapture } from "./capture-hooks.js";
 import { PLUGIN_HOOKS } from "./codebuddy-plugin.js";
+import { MCP_TOOLSET_KEY, MCP_TOOLSET_VERSION } from "./mcp-toolset.js";
 import {
   asInstalled,
   expectedPluginHooks,
@@ -9,6 +10,7 @@ import {
   type RecallHostLayout,
   type RecallInstallPlan,
   type RecallRuntime,
+  readInstalledRecallCommand,
   renderSkillContent,
   SKILL_VERSION,
 } from "./recall-installer.js";
@@ -99,7 +101,23 @@ export async function checkSkillDrift(
   runtime: RecallRuntime,
 ): Promise<SkillDrift | null> {
   const skill = await checkSkillFileDrift(layout, runtime);
-  if (skill === null || skill.state === "modified") return skill;
+  if (skill === null) {
+    // A host with no skill mechanism still has an entry whose tool set can
+    // age. One whose skill is merely missing is left alone: syncing would put
+    // back a file somebody removed.
+    if (skillFilePath(layout) !== null) return null;
+    return (await serverToolsetDrift(layout)) === "stale"
+      ? {
+          hostId: layout.hostId,
+          state: "stale",
+          installedVersion: null,
+          currentVersion: SKILL_VERSION,
+          skillPath: null,
+          detail: TOOLSET_STALE_DETAIL,
+        }
+      : null;
+  }
+  if (skill.state === "modified") return skill;
   const hooks = await pluginHooksDrift(layout, runtime);
   if (hooks === "modified") {
     return {
@@ -108,15 +126,19 @@ export async function checkSkillDrift(
       detail: "the plugin hook file was edited after install",
     };
   }
-  if (hooks === "stale" && skill.state === "in_sync") {
+  if (skill.state === "stale") return skill;
+  if (hooks === "stale") {
     return {
       ...skill,
       state: "stale",
       detail: "plugin hooks are still ours, but have moved since install",
     };
   }
+  if ((await serverToolsetDrift(layout)) === "stale") {
+    return { ...skill, state: "stale", detail: TOOLSET_STALE_DETAIL };
+  }
   const capture = await captureHooksDrift(layout);
-  if (capture === "stale" && skill.state === "in_sync") {
+  if (capture === "stale") {
     return {
       ...skill,
       state: "stale",
@@ -126,6 +148,29 @@ export async function checkSkillDrift(
   return skill;
 }
 
+const TOOLSET_STALE_DETAIL =
+  "the MCP entry predates the tool set this MemHub serves";
+
+/** Whether our MCP entry names the tool set this release serves.
+ *
+ * The entry is otherwise identical across upgrades, so a host that caches a
+ * server's tool list under a fingerprint of its config (WorkBuddy does) keeps
+ * listing tools that are gone until the entry itself changes. Only an entry
+ * marked as ours is read; a CLI-channel host has no entry at all.
+ */
+async function serverToolsetDrift(
+  layout: RecallHostLayout,
+): Promise<"in_sync" | "stale" | null> {
+  if (layout.channel !== "mcp") return null;
+  // A config the host cannot parse either is not ours to judge here; the
+  // install and link checks report it.
+  const installed = await readInstalledRecallCommand(layout).catch(() => null);
+  if (installed === null) return null;
+  return installed.env[MCP_TOOLSET_KEY] === MCP_TOOLSET_VERSION
+    ? "in_sync"
+    : "stale";
+}
+
 /** Detect duplicated command hooks left by earlier installs.
  *
  * `checkSkillDrift` historically compared only the Skill and plugin hooks;
diff --git a/memhub/packages/i18n/src/gui-i18n-current.ts b/memhub/packages/i18n/src/gui-i18n-current.ts
index 489a093..6eea397 100644
--- a/memhub/packages/i18n/src/gui-i18n-current.ts
+++ b/memhub/packages/i18n/src/gui-i18n-current.ts
@@ -108,6 +108,7 @@ export const currentChineseMessages = {
   "selectedMemory.version": "第 {revision} 版",
   "selectedMemory.sourceLine": "来源 {source} · {time}",
   "selectedMemory.timestamps": "对话时间 {conversation} · 沉淀时间 {created}",
+  "selectedMemory.writeTime": "写入时间 {created}",
   "selectedMemory.edit": "编辑",
   "selectedMemory.evidence": "原始依据",
   "selectedMemory.evidenceLine": "{author} · 原对话 {time}：",
@@ -258,6 +259,7 @@ export const currentEnglishMessages: Record<
   "selectedMemory.sourceLine": "Source {source} · {time}",
   "selectedMemory.timestamps":
     "Conversation time {conversation} · Created {created}",
+  "selectedMemory.writeTime": "Written {created}",
   "selectedMemory.edit": "Edit",
   "selectedMemory.evidence": "Original evidence",
   "selectedMemory.evidenceLine": "{author} · Original conversation {time}:",
@@ -418,6 +420,7 @@ export const currentSpanishMessages: Record<
   "selectedMemory.sourceLine": "Origen {source} · {time}",
   "selectedMemory.timestamps":
     "Conversación {conversation} · Guardada {created}",
+  "selectedMemory.writeTime": "Escrita {created}",
   "selectedMemory.edit": "Editar",
   "selectedMemory.evidence": "Evidencia original",
   "selectedMemory.evidenceLine": "{author} · Conversación original {time}:",
diff --git a/memhub/packages/i18n/src/gui-i18n-english.ts b/memhub/packages/i18n/src/gui-i18n-english.ts
index caff69a..1559d70 100644
--- a/memhub/packages/i18n/src/gui-i18n-english.ts
+++ b/memhub/packages/i18n/src/gui-i18n-english.ts
@@ -1782,6 +1782,7 @@ const baseEnglishMessages = {
     " · Import count is still updating; refresh to check",
   "memory:对话于 {day}": "Conversation on {day}",
   "memory:导入于 {day}": "Imported on {day}",
+  "memory:写入于 {day}": "Written on {day}",
   "memory:已开始处理所选对话": "Processing selected conversations",
   "memory:处理请求已受理": "Processing request received",
   "memory:模型连接已通过。首批记忆生成后，会逐步出现在记忆列表里。":
diff --git a/memhub/packages/i18n/src/gui-i18n-memory.ts b/memhub/packages/i18n/src/gui-i18n-memory.ts
index d27ee09..9db1fb3 100644
--- a/memhub/packages/i18n/src/gui-i18n-memory.ts
+++ b/memhub/packages/i18n/src/gui-i18n-memory.ts
@@ -336,6 +336,7 @@ export const memoryChineseMessages = {
   "memory: · 导入数量仍在更新，可刷新核对": " · 导入数量仍在更新，可刷新核对",
   "memory:对话于 {day}": "对话于 {day}",
   "memory:导入于 {day}": "导入于 {day}",
+  "memory:写入于 {day}": "写入于 {day}",
   "memory:已开始处理所选对话": "已开始处理所选对话",
   "memory:处理请求已受理": "处理请求已受理",
   "memory:模型连接已通过。首批记忆生成后，会逐步出现在记忆列表里。":
@@ -750,6 +751,7 @@ export const memorySpanishMessages: Record<
     " · La importación continúa; actualiza para verificar",
   "memory:对话于 {day}": "Conversación del {day}",
   "memory:导入于 {day}": "Importada el {day}",
+  "memory:写入于 {day}": "Escrita el {day}",
   "memory:已开始处理所选对话": "Procesamiento de conversaciones iniciado",
   "memory:处理请求已受理": "Solicitud de procesamiento aceptada",
   "memory:模型连接已通过。首批记忆生成后，会逐步出现在记忆列表里。":
diff --git a/memhub/packages/runtime-control/src/host-permissions.ts b/memhub/packages/runtime-control/src/host-permissions.ts
index 67d2a54..9f42f93 100644
--- a/memhub/packages/runtime-control/src/host-permissions.ts
+++ b/memhub/packages/runtime-control/src/host-permissions.ts
@@ -41,6 +41,7 @@ import {
   repairInstalledRecallCommands,
   verifyRecallInstall,
 } from "@one-memory/host-bridge/installer";
+import type { DiagnosticInput, DiagnosticSink } from "./diagnostic-log.js";
 import {
   hasSyncHookArtifacts,
   loadSyncHookAssets,
@@ -96,8 +97,22 @@ interface InstallationBinding {
 }
 
 const CREDENTIAL_ENV = "ONE_MEMORY_HOST_CREDENTIAL_FILE";
+/** The build whose host refresh last completed, beside the grants. */
+const HOST_REFRESH_FILE = "host-refresh.json";
 const locks = new Map<string, Promise<void>>();
 
+/** What one pass of `refreshInstalledHosts` did, by host id. */
+export interface HostRefreshReport {
+  /** `already_refreshed` when this build had already completed a pass. */
+  state: "refreshed" | "already_refreshed";
+  checked: number;
+  /** Hosts whose installed command or discovery path was moved. */
+  repointed: string[];
+  /** Hosts whose stale skill, hooks, plugin files or MCP entry were rewritten. */
+  synced: string[];
+  failed: string[];
+}
+
 export const LOCAL_EXPERIENCE_UNAVAILABLE = "体验版暂不支持";
 
 /** Trusted local Runtime control plane. Never exposed through MCP or the Core API.
@@ -201,11 +216,20 @@ export class HostPermissions {
    * touching authorization. A skill the user edited is left for the detail page.
    */
   async syncStaleInstalls(layouts: RecallHostLayout[]): Promise<string[]> {
+    return (await this.syncStale(layouts)).synced;
+  }
+
+  private async syncStale(
+    layouts: RecallHostLayout[],
+  ): Promise<{ synced: string[]; failed: string[] }> {
     const synced: string[] = [];
+    const failed: string[] = [];
     for (const layout of layouts) {
       try {
         await this.withHostLock(layout, async () => {
-          await this.requireLinked(layout);
+          // Not linked from this data directory is not ours to sync, and not
+          // a failure either.
+          if (!(await this.isLinked(layout))) return;
           const sync = await planSkillSync(layout, this.runtimeFor(layout));
           if (sync === null || sync.requiresConfirmation) return;
           const preview = await previewRecallInstall(sync.plan);
@@ -215,9 +239,128 @@ export class HostPermissions {
         });
       } catch {
         // One host that cannot be synced must not stop the runtime starting.
+        failed.push(layout.hostId);
       }
     }
-    return synced;
+    return { synced, failed };
+  }
+
+  /** Refresh every host this directory installed, once per build.
+   *
+   * An upgrade changes the call rules, hooks and MCP entry we would write, but
+   * nothing rewrote the copies hosts already had, so they kept telling the
+   * model to call tools that were gone. This runs the same repoint and stale
+   * sync the detail page offers, without asking, for files that are still
+   * exactly ours; an edited skill stays for the user to confirm.
+   *
+   * The build it last completed for is recorded beside the grants, so later
+   * launches of the same build cost one file read. A run in which any host
+   * failed is not recorded, so the next launch tries again. With no build to
+   * name (a source checkout) it runs every time and records nothing.
+   */
+  async refreshInstalledHosts(
+    layouts: RecallHostLayout[],
+    options: { build?: string; diagnostics?: DiagnosticSink } = {},
+  ): Promise<HostRefreshReport> {
+    const log = (input: DiagnosticInput) =>
+      options.diagnostics?.log(input).catch(() => undefined);
+    const build = options.build?.trim() || undefined;
+    if (build !== undefined && (await this.readRefreshedBuild()) === build) {
+      await log({
+        level: "info",
+        component: "host.refresh",
+        event: "host_refresh.skipped",
+        outcome: "skipped",
+        details: { reason: "already_refreshed", count: layouts.length },
+      });
+      return {
+        state: "already_refreshed",
+        checked: layouts.length,
+        repointed: [],
+        synced: [],
+        failed: [],
+      };
+    }
+    const started = Date.now();
+    const repointed: string[] = [];
+    const failed: string[] = [];
+    for (const layout of layouts) {
+      try {
+        repointed.push(...(await this.repairInstalledCommands([layout])));
+      } catch {
+        failed.push(layout.hostId);
+      }
+    }
+    const stale = await this.syncStale(layouts);
+    failed.push(...stale.failed.filter((hostId) => !failed.includes(hostId)));
+    for (const hostId of new Set([...repointed, ...stale.synced])) {
+      await log({
+        level: "info",
+        component: "host.refresh",
+        event: "host_refresh.host_updated",
+        outcome: "succeeded",
+        details: { source_id: hostId },
+      });
+    }
+    for (const hostId of failed) {
+      await log({
+        level: "warn",
+        component: "host.refresh",
+        event: "host_refresh.host_failed",
+        outcome: "failed",
+        error_code: "host_refresh_failed",
+        details: { source_id: hostId },
+      });
+    }
+    if (build !== undefined && failed.length === 0) {
+      await this.writeRefreshedBuild(build).catch(() => undefined);
+    }
+    await log({
+      level: failed.length === 0 ? "info" : "warn",
+      component: "host.refresh",
+      event: "host_refresh.completed",
+      outcome: failed.length === 0 ? "succeeded" : "degraded",
+      duration_ms: Date.now() - started,
+      details: {
+        count: layouts.length,
+        updated_count: new Set([...repointed, ...stale.synced]).size,
+        error_count: failed.length,
+      },
+    });
+    return {
+      state: "refreshed",
+      checked: layouts.length,
+      repointed,
+      synced: stale.synced,
+      failed,
+    };
+  }
+
+  private async readRefreshedBuild(): Promise<string | null> {
+    try {
+      return await storageGuard(async () => {
+        const file = path.join(this.directory, HOST_REFRESH_FILE);
+        if ((await optionalStat(file)) === null) return null;
+        const value: unknown = JSON.parse(await privateRead(file));
+        return isObject(value) &&
+          value.schemaVersion === 1 &&
+          typeof value.build === "string"
+          ? value.build
+          : null;
+      });
+    } catch {
+      // Unreadable reads as never refreshed: the refresh is idempotent.
+      return null;
+    }
+  }
+
+  private async writeRefreshedBuild(build: string): Promise<void> {
+    await storageGuard(() =>
+      privateWrite(
+        path.join(this.directory, HOST_REFRESH_FILE),
+        `${JSON.stringify({ schemaVersion: 1, build })}\n`,
+      ),
+    );
   }
 
   /** Same repair for one host. Does not revoke or reissue its authorization. */
@@ -942,6 +1085,16 @@ export class HostPermissions {
       throw new Error("host_permission_installation_incomplete");
   }
 
+  /** `requireLinked` as a question, for sweeps that skip hosts not ours. */
+  private async isLinked(layout: RecallHostLayout): Promise<boolean> {
+    if (!supported(layout)) return false;
+    const current = await this.inspect(layout, false);
+    return (
+      (current.state === "approved" || current.state === "install-only") &&
+      current.reason === null
+    );
+  }
+
   /** Turning off a host with nothing of ours on it is an empty removal, not a
    * failure; one whose state cannot be settled is an empty removal carrying
    * what to look at, which uninstall refuses. Only what is ours for certain
diff --git a/memhub/packages/runtime-control/src/index.ts b/memhub/packages/runtime-control/src/index.ts
index 285875a..7f7e2e1 100644
--- a/memhub/packages/runtime-control/src/index.ts
+++ b/memhub/packages/runtime-control/src/index.ts
@@ -79,6 +79,7 @@ export {
 } from "./host-control-client.js";
 export {
   HostPermissions,
+  type HostRefreshReport,
   LOCAL_EXPERIENCE_UNAVAILABLE,
   type RecallPermissionStatus,
   type RevokeRecallPermissionInput,
diff --git a/memhub/packages/runtime-control/src/runtime-cli.test.ts b/memhub/packages/runtime-control/src/runtime-cli.test.ts
index 175b111..8f724b6 100644
--- a/memhub/packages/runtime-control/src/runtime-cli.test.ts
+++ b/memhub/packages/runtime-control/src/runtime-cli.test.ts
@@ -226,6 +226,7 @@ describe("host runtime attestation", () => {
     "repairInstalledCommand",
     "repairInstalledCommands",
     "syncStaleInstalls",
+    "refreshInstalledHosts",
     "previewInstall",
     "install",
     "setPermission",
diff --git a/memhub/packages/runtime-control/src/runtime-cli.ts b/memhub/packages/runtime-control/src/runtime-cli.ts
index ad3d207..70e906b 100644
--- a/memhub/packages/runtime-control/src/runtime-cli.ts
+++ b/memhub/packages/runtime-control/src/runtime-cli.ts
@@ -600,6 +600,7 @@ export function guardedPermissions(
     "previewSyncHook",
     "applySyncHook",
     "syncStaleInstalls",
+    "refreshInstalledHosts",
   ]);
   return new Proxy(owner, {
     get(target, property, receiver) {
@@ -739,10 +740,17 @@ export async function createRuntimeService(
       ).filter((host) => host.hostId !== "openclaw");
       await verifyHostRuntime().then(
         async () => {
-          await permissions.repairInstalledCommands(
+          // Every Runtime start after an upgrade, the App's or `memhub
+          // upgrade`'s, comes through here first; once per build after that.
+          await permissions.refreshInstalledHosts(
             hosts.map((host) => host.layout),
+            {
+              build:
+                config.buildIdentity?.buildNumber ??
+                config.buildIdentity?.version,
+              diagnostics: hostDiagnostics,
+            },
           );
-          await permissions.syncStaleInstalls(hosts.map((host) => host.layout));
         },
         () => undefined,
       );
diff --git a/memhub/packages/runtime-control/src/runtime-service.test.ts b/memhub/packages/runtime-control/src/runtime-service.test.ts
index 382c206..f4bcad1 100644
--- a/memhub/packages/runtime-control/src/runtime-service.test.ts
+++ b/memhub/packages/runtime-control/src/runtime-service.test.ts
@@ -2,9 +2,11 @@ import {
   chmod,
   mkdir,
   mkdtemp,
+  readdir,
   readFile,
   rm,
   stat,
+  symlink,
   writeFile,
 } from "node:fs/promises";
 import { tmpdir } from "node:os";
@@ -38,6 +40,38 @@ function supervisor() {
   };
 }
 
+const PREVIOUS_BOOT = "00000000-0000-4000-8000-00000000000a";
+const CURRENT_BOOT = "00000000-0000-4000-8000-00000000000b";
+
+/** Seed a lock as an existing owner would leave it; no owner = legacy debris. */
+async function seedLock(dataDir: string, owner?: object): Promise<string> {
+  const lock = path.join(dataDir, "runtime-service.lock");
+  await mkdir(lock, { mode: 0o700 });
+  if (owner) {
+    await writeFile(path.join(lock, "owner.json"), JSON.stringify(owner), {
+      mode: 0o600,
+    });
+  }
+  await chmod(lock, 0o700);
+  return lock;
+}
+
+async function ownerOf(dataDir: string): Promise<Record<string, unknown>> {
+  return JSON.parse(
+    await readFile(
+      path.join(dataDir, "runtime-service.lock", "owner.json"),
+      "utf8",
+    ),
+  );
+}
+
+/** Every lock, staging and cleared-lease entry in the data dir. */
+async function lockEntries(dataDir: string): Promise<string[]> {
+  return (await readdir(dataDir))
+    .filter((name) => name.startsWith("runtime-service.lock"))
+    .sort();
+}
+
 describe("RuntimeService", () => {
   it("publishes and removes an outer Runtime receipt without secrets", async () => {
     const dataDir = await directory("runtime-service-");
@@ -71,18 +105,7 @@ describe("RuntimeService", () => {
 
   it("refuses a live owner of the same Runtime home", async () => {
     const dataDir = await directory("runtime-service-live-");
-    const lock = path.join(dataDir, "runtime-service.lock");
-    await import("node:fs/promises").then(({ mkdir }) =>
-      mkdir(lock, { mode: 0o700 }),
-    );
-    await writeFile(
-      path.join(lock, "owner.json"),
-      JSON.stringify({ pid: 777 }),
-      {
-        mode: 0o600,
-      },
-    );
-    await chmod(lock, 0o700);
+    await seedLock(dataDir, { pid: 777 });
     const service = new RuntimeService({
       dataDir,
       createSupervisor: supervisor,
@@ -94,27 +117,228 @@ describe("RuntimeService", () => {
     );
   });
 
-  it("reclaims a stale validated lease", async () => {
-    const dataDir = await directory("runtime-service-stale-");
-    const lock = path.join(dataDir, "runtime-service.lock");
-    await import("node:fs/promises").then(({ mkdir }) =>
-      mkdir(lock, { mode: 0o700 }),
-    );
-    await writeFile(
-      path.join(lock, "owner.json"),
-      JSON.stringify({ pid: 777 }),
-      {
-        mode: 0o600,
-      },
-    );
-    await chmod(lock, 0o700);
+  it("reclaims a previous-boot lease even when a recycled PID answers the liveness probe", async () => {
+    // Issue #184: a forced power-off leaves runtime-service.lock behind;
+    // after the reboot the owner's PID belongs to another process, so a
+    // liveness-only check reports "already running" forever and launchd's
+    // KeepAlive loops the wrapper. A lease from another boot can never have
+    // a living owner, whatever the PID table says.
+    const dataDir = await directory("runtime-service-prevboot-");
+    await seedLock(dataDir, { pid: 777, boot_id: PREVIOUS_BOOT });
     const service = new RuntimeService({
       dataDir,
       createSupervisor: supervisor,
+      pid: 4321,
+      processIsAlive: () => true,
+      bootId: () => CURRENT_BOOT,
+    });
+
+    await service.start();
+    expect(await ownerOf(dataDir)).toMatchObject({
+      pid: 4321,
+      boot_id: CURRENT_BOOT,
+    });
+    await service.stop();
+    expect(await lockEntries(dataDir)).toEqual([]);
+  });
+
+  it("refuses a live owner from the current boot", async () => {
+    const dataDir = await directory("runtime-service-sameboot-live-");
+    await seedLock(dataDir, { pid: 777, boot_id: CURRENT_BOOT });
+    const service = new RuntimeService({
+      dataDir,
+      createSupervisor: supervisor,
+      pid: 4321,
+      processIsAlive: (pid) => pid === 777,
+      bootId: () => CURRENT_BOOT,
+    });
+
+    await expect(service.start()).rejects.toThrow(
+      "runtime_service_already_running",
+    );
+    expect(await ownerOf(dataDir)).toMatchObject({ pid: 777 });
+  });
+
+  it("reclaims a dead owner from the current boot", async () => {
+    const dataDir = await directory("runtime-service-sameboot-dead-");
+    await seedLock(dataDir, { pid: 777, boot_id: CURRENT_BOOT });
+    const service = new RuntimeService({
+      dataDir,
+      createSupervisor: supervisor,
+      pid: 4321,
+      processIsAlive: () => false,
+      bootId: () => CURRENT_BOOT,
+    });
+
+    await service.start();
+    expect(await ownerOf(dataDir)).toMatchObject({ pid: 4321 });
+    await service.stop();
+  });
+
+  it.each([
+    ["the lease predates boot ids", { pid: 777 }, CURRENT_BOOT],
+    ["the lease has an empty boot id", { pid: 777, boot_id: "" }, CURRENT_BOOT],
+    [
+      "the current boot id is unknown",
+      { pid: 777, boot_id: PREVIOUS_BOOT },
+      null,
+    ],
+  ])("falls back to the PID probe when %s", async (_case, owner, current) => {
+    const dataDir = await directory("runtime-service-unknown-boot-");
+    await seedLock(dataDir, owner);
+    const live = new RuntimeService({
+      dataDir,
+      createSupervisor: supervisor,
+      pid: 4321,
+      processIsAlive: (pid) => pid === 777,
+      bootId: () => current,
+    });
+    await expect(live.start()).rejects.toThrow(
+      "runtime_service_already_running",
+    );
+    expect(await ownerOf(dataDir)).toMatchObject({ pid: 777 });
+
+    const dead = new RuntimeService({
+      dataDir,
+      createSupervisor: supervisor,
+      pid: 4321,
+      processIsAlive: () => false,
+      bootId: () => current,
+    });
+    await dead.start();
+    expect(await ownerOf(dataDir)).toMatchObject({ pid: 4321 });
+    await dead.stop();
+  });
+
+  it("records the platform boot id in owner.json", async () => {
+    const dataDir = await directory("runtime-service-boot-id-");
+    const service = new RuntimeService({
+      dataDir,
+      createSupervisor: supervisor,
+    });
+
+    await service.start();
+    const owner = await ownerOf(dataDir);
+    if (process.platform === "darwin" || process.platform === "linux") {
+      expect(owner.boot_id).toMatch(/^[0-9A-Za-z-]{8,64}$/);
+    } else {
+      expect(owner.boot_id).toBeNull();
+    }
+    expect(owner.pid).toBe(process.pid);
+    await service.stop();
+  });
+
+  it("leaves no lock behind when publishing fails before the rename", async () => {
+    // The lock only appears through a rename of a fully written staging
+    // directory, so a failure (or power cut) while staging cannot leave a
+    // lock without owner.json.
+    const dataDir = await directory("runtime-service-publish-failure-");
+    const broken = new RuntimeService({
+      dataDir,
+      createSupervisor: supervisor,
+      bootId: () => {
+        throw new Error("simulated crash while staging");
+      },
+    });
+
+    await expect(broken.start()).rejects.toThrow(
+      "runtime_service_lock_unavailable",
+    );
+    expect(await lockEntries(dataDir)).toEqual([]);
+
+    const next = new RuntimeService({
+      dataDir,
+      createSupervisor: supervisor,
+      pid: 4321,
+      bootId: () => CURRENT_BOOT,
+    });
+    await next.start();
+    expect(await ownerOf(dataDir)).toMatchObject({ pid: 4321 });
+    await next.stop();
+  });
+
+  it("clears an owner-less legacy lock created before the current boot", async () => {
+    // Older versions created the lock with mkdir and wrote owner.json
+    // afterwards; a power cut in between left an empty lock that used to
+    // fail as runtime_service_lock_invalid forever.
+    const dataDir = await directory("runtime-service-legacy-prevboot-");
+    const lock = await seedLock(dataDir);
+    const createdAt = (await stat(lock)).mtimeMs;
+    const service = new RuntimeService({
+      dataDir,
+      createSupervisor: supervisor,
+      pid: 4321,
+      bootId: () => CURRENT_BOOT,
+      legacyBootTime: () => createdAt + 60_000,
+    });
+
+    await service.start();
+    expect(await ownerOf(dataDir)).toMatchObject({ pid: 4321 });
+    await service.stop();
+  });
+
+  it.each([
+    ["it may be from the current boot", (createdAt: number) => createdAt - 1],
+    ["the boot time is unknown", () => null],
+  ])(
+    "keeps refusing an owner-less legacy lock when %s",
+    async (_case, boot) => {
+      const dataDir = await directory("runtime-service-legacy-sameboot-");
+      const lock = await seedLock(dataDir);
+      const createdAt = (await stat(lock)).mtimeMs;
+      const service = new RuntimeService({
+        dataDir,
+        createSupervisor: supervisor,
+        bootId: () => CURRENT_BOOT,
+        legacyBootTime: () => boot(createdAt),
+      });
+
+      await expect(service.start()).rejects.toThrow(
+        "runtime_service_lock_invalid",
+      );
+      // rename() may replace an empty directory; the publish step must not.
+      expect(await readdir(lock)).toEqual([]);
+      expect(await lockEntries(dataDir)).toEqual(["runtime-service.lock"]);
+    },
+  );
+
+  it("removes staging and cleared-lease debris after acquiring", async () => {
+    const dataDir = await directory("runtime-service-debris-");
+    const staging = path.join(dataDir, "runtime-service.lock.tmp-crashed");
+    await mkdir(staging, { mode: 0o700 });
+    await writeFile(
+      path.join(staging, "owner.json"),
+      JSON.stringify({ pid: 9 }),
+    );
+    await mkdir(path.join(dataDir, "runtime-service.lock.stale.crashed"), {
+      mode: 0o700,
+    });
+    const service = new RuntimeService({
+      dataDir,
+      createSupervisor: supervisor,
+      pid: 4321,
+      bootId: () => CURRENT_BOOT,
+    });
+
+    await service.start();
+    expect(await lockEntries(dataDir)).toEqual(["runtime-service.lock"]);
+    expect(await ownerOf(dataDir)).toMatchObject({ pid: 4321 });
+    await service.stop();
+    expect(await lockEntries(dataDir)).toEqual([]);
+  });
+
+  it("reclaims a stale validated lease", async () => {
+    const dataDir = await directory("runtime-service-stale-");
+    await seedLock(dataDir, { pid: 777 });
+    const service = new RuntimeService({
+      dataDir,
+      createSupervisor: supervisor,
+      pid: 4321,
       processIsAlive: () => false,
     });
 
     await service.start();
+    expect(await ownerOf(dataDir)).toMatchObject({ pid: 4321 });
     await service.stop();
   });
 
@@ -386,7 +610,7 @@ describe("RuntimeService", () => {
     const real = path.join(parent, "real");
     const link = path.join(parent, "data");
     await mkdir(real, { mode: 0o755 });
-    await import("node:fs/promises").then(({ symlink }) => symlink(real, link));
+    await symlink(real, link);
     const service = new RuntimeService({
       dataDir: link,
       createSupervisor: () => supervisor(),
diff --git a/memhub/packages/runtime-control/src/runtime-service.ts b/memhub/packages/runtime-control/src/runtime-service.ts
index d3bdc1d..86432e3 100644
--- a/memhub/packages/runtime-control/src/runtime-service.ts
+++ b/memhub/packages/runtime-control/src/runtime-service.ts
@@ -1,12 +1,14 @@
+import { execFileSync } from "node:child_process";
 import { randomUUID } from "node:crypto";
+import { readFileSync, type Stats } from "node:fs";
 import {
   chmod,
   lstat,
   mkdir,
+  readdir,
   readFile,
   rename,
   rm,
-  rmdir,
   writeFile,
 } from "node:fs/promises";
 import path from "node:path";
@@ -35,6 +37,14 @@ export interface RuntimeServiceOptions {
   ): ManagedRuntime | Promise<ManagedRuntime>;
   pid?: number;
   processIsAlive?: (pid: number) => boolean;
+  /** Identity of the current boot (`kern.bootsessionuuid` on macOS,
+   * `/proc/sys/kernel/random/boot_id` on Linux, `null` when unknown).
+   * Injectable so tests can simulate a reboot without one. */
+  bootId?: () => string | null;
+  /** Wall-clock epoch (ms) the kernel reports for the current boot, or `null`
+   * when unknown. Consulted only for a legacy lock directory without
+   * owner.json; see removeUnownedLease. Injectable for tests. */
+  legacyBootTime?: () => number | null;
   diagnostics?: DiagnosticSink;
 }
 
@@ -186,29 +196,81 @@ export class RuntimeService {
     const lock = this.lockPath();
     for (let attempt = 0; attempt < 2; attempt += 1) {
       try {
-        await mkdir(lock, { mode: 0o700 });
-        const identity = await lstat(lock);
-        this.lease = { dev: identity.dev, ino: identity.ino };
-        try {
-          await privateWrite(
-            path.join(lock, "owner.json"),
-            JSON.stringify({
-              pid: this.pid(),
-              created_at: new Date().toISOString(),
-            }),
-          );
-        } catch (error) {
-          await this.releaseLease();
-          throw error;
-        }
-        return;
+        await this.publishLease(lock);
+        break;
       } catch (error) {
-        if ((error as NodeJS.ErrnoException).code !== "EEXIST" || attempt > 0) {
+        if (!leaseTaken(error) || attempt > 0) {
           throw new Error("runtime_service_lock_unavailable", { cause: error });
         }
         await this.removeStaleLease(lock);
       }
     }
+    await this.sweepLeaseDebris();
+  }
+
+  /** Publish the lease in one step: the lock directory becomes visible with
+   * owner.json (pid, created_at, boot_id) already inside it, so a power cut
+   * can leave a staging directory behind but never a lock without an owner.
+   * The old mkdir -> write sequence had exactly that window (issue #184). */
+  private async publishLease(lock: string): Promise<void> {
+    const staging = `${lock}${STAGING_INFIX}${randomUUID()}`;
+    await mkdir(staging, { mode: 0o700 });
+    try {
+      await privateWrite(
+        path.join(staging, "owner.json"),
+        JSON.stringify({
+          pid: this.pid(),
+          created_at: new Date().toISOString(),
+          boot_id: this.currentBootId(),
+        }),
+      );
+      // POSIX rename() silently REPLACES an existing *empty* directory, so on
+      // its own it is not an exclusive create. A non-empty lock (every lease
+      // this version publishes) makes it fail with EEXIST/ENOTEMPTY. An empty
+      // lock can only come from an older version's two-step mkdir -> write:
+      // either power-cut debris, or a live acquire in flight. The latter must
+      // not be overwritten, and the former has to go through
+      // removeStaleLease's legacy rule rather than be replaced blindly, so an
+      // existing lock of any kind is reported as taken here. The residual race
+      // (an old binary mkdir-ing between this check and the rename) needs an
+      // old and a new Runtime starting on the same home within microseconds;
+      // this version never creates an empty lock, since releaseLease also
+      // removes it with a single rename.
+      if (await pathExists(lock)) throw leaseTakenError();
+      await rename(staging, lock);
+    } catch (error) {
+      await rm(staging, { recursive: true, force: true }).catch(
+        () => undefined,
+      );
+      throw error;
+    }
+    // rename() keeps the inode, so this is the published directory's identity;
+    // releaseLease compares it before touching the lock.
+    const identity = await lstat(lock);
+    this.lease = { dev: identity.dev, ino: identity.ino };
+  }
+
+  /** Remove staging and cleared-lease directories left behind by a crash.
+   * Runs only while this process holds the lease, so a staging directory is
+   * either debris or belongs to a competitor that has already lost (its
+   * rename cannot succeed over our non-empty lock). Best-effort. */
+  private async sweepLeaseDebris(): Promise<void> {
+    const base = path.basename(this.lockPath());
+    const names = await readdir(this.options.dataDir).catch((): string[] => []);
+    await Promise.all(
+      names
+        .filter(
+          (name) =>
+            name.startsWith(`${base}${STAGING_INFIX}`) ||
+            name.startsWith(`${base}${STALE_INFIX}`),
+        )
+        .map((name) =>
+          rm(path.join(this.options.dataDir, name), {
+            recursive: true,
+            force: true,
+          }).catch(() => undefined),
+        ),
+    );
   }
 
   private async removeStaleLease(lock: string): Promise<void> {
@@ -223,25 +285,81 @@ export class RuntimeService {
     ) {
       throw new Error("runtime_service_lock_unsafe");
     }
+    let raw: string;
+    try {
+      raw = await readFile(path.join(lock, "owner.json"), "utf8");
+    } catch (error) {
+      if ((error as NodeJS.ErrnoException).code === "ENOENT") {
+        await this.removeUnownedLease(lock, info);
+        return;
+      }
+      throw new Error("runtime_service_lock_invalid", { cause: error });
+    }
     let owner: unknown;
     try {
-      owner = JSON.parse(await readFile(path.join(lock, "owner.json"), "utf8"));
+      owner = JSON.parse(raw);
     } catch {
       throw new Error("runtime_service_lock_invalid");
     }
-    const pid =
-      owner && typeof owner === "object"
-        ? (owner as { pid?: unknown }).pid
-        : undefined;
+    if (!owner || typeof owner !== "object" || Array.isArray(owner)) {
+      throw new Error("runtime_service_lock_invalid");
+    }
+    // A lease written during another boot cannot have a living owner: every
+    // process of that boot is gone, and a PID that answers the liveness probe
+    // now merely reused the number. Without this, a forced power-off left the
+    // lock behind, the reboot recycled the owner's PID, and launchd's
+    // KeepAlive restarted the wrapper into runtime_service_already_running
+    // without end (issue #184). The kernel's boot identity is compared, not
+    // the wall clock: the clock can jump forward while a lease is held (no
+    // RTC, VM resume, NTP or manual correction) and make a live lease look
+    // older than the boot. When either side is unknown (a lease from an older
+    // version, an unsupported platform) the PID probe below decides, as before.
+    const ownerBoot = (owner as { boot_id?: unknown }).boot_id;
+    const currentBoot = this.currentBootId();
+    if (
+      typeof ownerBoot === "string" &&
+      ownerBoot !== "" &&
+      currentBoot !== null &&
+      ownerBoot !== currentBoot
+    ) {
+      await this.clearLeaseDir(lock);
+      return;
+    }
+    const pid = (owner as { pid?: unknown }).pid;
     if (typeof pid !== "number" || !Number.isInteger(pid) || pid <= 1) {
       throw new Error("runtime_service_lock_invalid");
     }
     if ((this.options.processIsAlive ?? processIsAlive)(pid)) {
       throw new Error("runtime_service_already_running");
     }
-    const stale = `${lock}.stale.${randomUUID()}`;
+    await this.clearLeaseDir(lock);
+  }
+
+  /** A lock directory without owner.json. This version publishes owner.json
+   * together with the lock and releases it with one rename, so such a
+   * directory was written by an older version (mkdir -> write, where a power
+   * cut in between leaves it empty) or by hand. It names no PID and no boot,
+   * so the only usable evidence is "created before the current boot": clear
+   * it then, and otherwise keep the previous refusal
+   * (runtime_service_lock_invalid), including when the boot time is unknown.
+   *
+   * The kernel boot time and the directory mtime are both wall-clock values,
+   * which is why lease ownership itself no longer depends on them. Here a
+   * forward clock jump can only hurt if an older binary is live inside its
+   * mkdir -> write window at that very moment; a backward jump only keeps the
+   * old refusal. */
+  private async removeUnownedLease(lock: string, info: Stats): Promise<void> {
+    const bootTime = (this.options.legacyBootTime ?? currentBootTimeMs)();
+    if (bootTime === null || !(info.mtimeMs < bootTime)) {
+      throw new Error("runtime_service_lock_invalid");
+    }
+    await this.clearLeaseDir(lock);
+  }
+
+  private async clearLeaseDir(lock: string): Promise<void> {
+    const stale = `${lock}${STALE_INFIX}${randomUUID()}`;
     await rename(lock, stale);
-    await rm(stale, { recursive: true });
+    await rm(stale, { recursive: true, force: true });
   }
 
   private async releaseLease(): Promise<void> {
@@ -252,8 +370,19 @@ export class RuntimeService {
     const current = await lstat(lock).catch(() => null);
     if (!current || current.dev !== lease.dev || current.ino !== lease.ino)
       return;
-    await rm(path.join(lock, "owner.json"), { force: true });
-    await rmdir(lock).catch(() => undefined);
+    // Take the lock away with one rename so no observer ever sees it without
+    // owner.json: an owner-less lock is left to mean legacy debris.
+    const released = `${lock}${STALE_INFIX}${randomUUID()}`;
+    try {
+      await rename(lock, released);
+    } catch {
+      return;
+    }
+    await rm(released, { recursive: true, force: true }).catch(() => undefined);
+  }
+
+  private currentBootId(): string | null {
+    return (this.options.bootId ?? currentBootId)();
   }
 
   private async writeReceipt(
@@ -302,6 +431,91 @@ export class RuntimeService {
   }
 }
 
+const STAGING_INFIX = ".tmp-";
+const STALE_INFIX = ".stale.";
+
+function leaseTaken(error: unknown): boolean {
+  const code = (error as NodeJS.ErrnoException | undefined)?.code;
+  return code === "EEXIST" || code === "ENOTEMPTY";
+}
+
+function leaseTakenError(): NodeJS.ErrnoException {
+  return Object.assign(new Error("runtime_service_lock_exists"), {
+    code: "EEXIST",
+  });
+}
+
+async function pathExists(target: string): Promise<boolean> {
+  try {
+    await lstat(target);
+    return true;
+  } catch (error) {
+    if ((error as NodeJS.ErrnoException).code === "ENOENT") return false;
+    throw error;
+  }
+}
+
+function sysctl(name: string): string {
+  return execFileSync("/usr/sbin/sysctl", ["-n", name], {
+    encoding: "utf8",
+    timeout: 1000,
+    maxBuffer: 4096,
+    killSignal: "SIGKILL",
+    shell: false,
+    stdio: ["ignore", "pipe", "ignore"],
+    env: { PATH: "/usr/bin:/bin:/usr/sbin:/sbin", LANG: "C", LC_ALL: "C" },
+  });
+}
+
+let cachedBootId: string | null | undefined;
+
+/** The kernel's identity for the current boot, cached per process (it cannot
+ * change while the process lives). `null` on unsupported platforms or when
+ * the source cannot be read; the lease logic then falls back to the PID
+ * probe. */
+function currentBootId(): string | null {
+  if (cachedBootId === undefined) cachedBootId = readBootId();
+  return cachedBootId;
+}
+
+function readBootId(): string | null {
+  try {
+    let raw: string;
+    if (process.platform === "darwin") raw = sysctl("kern.bootsessionuuid");
+    else if (process.platform === "linux")
+      raw = readFileSync("/proc/sys/kernel/random/boot_id", "utf8");
+    else return null;
+    const id = raw.trim();
+    return /^[0-9A-Za-z-]{8,64}$/.test(id) ? id : null;
+  } catch {
+    return null;
+  }
+}
+
+/** Wall-clock epoch (ms) the kernel reports for the current boot, or `null`.
+ * Used only for owner-less legacy lock directories. */
+function currentBootTimeMs(): number | null {
+  try {
+    if (process.platform === "darwin") {
+      const match = /\bsec\s*=\s*(\d+)\s*,\s*usec\s*=\s*(\d+)/.exec(
+        sysctl("kern.boottime"),
+      );
+      return match
+        ? Number(match[1]) * 1000 + Math.floor(Number(match[2]) / 1000)
+        : null;
+    }
+    if (process.platform === "linux") {
+      const match = /^btime\s+(\d+)\s*$/m.exec(
+        readFileSync("/proc/stat", "utf8"),
+      );
+      return match ? Number(match[1]) * 1000 : null;
+    }
+  } catch {
+    // Unknown boot time: the caller keeps refusing.
+  }
+  return null;
+}
+
 async function privateWrite(target: string, value: string): Promise<void> {
   const temporary = `${target}.${randomUUID()}.tmp`;
   try {
diff --git a/memhub/packages/runtime-control/src/sdk-migration-log-order.test.ts b/memhub/packages/runtime-control/src/sdk-migration-log-order.test.ts
new file mode 100644
index 0000000..5e3a891
--- /dev/null
+++ b/memhub/packages/runtime-control/src/sdk-migration-log-order.test.ts
@@ -0,0 +1,101 @@
+import { mkdtemp, readFile, rm, writeFile } from "node:fs/promises";
+import os from "node:os";
+import path from "node:path";
+import { describe, expect, it } from "vitest";
+import {
+  sdkMigrationLogPath,
+  writeSdkMigrationLog,
+} from "./sdk-migration-log.js";
+
+async function events(dataDir: string): Promise<string[]> {
+  return (await readFile(sdkMigrationLogPath(dataDir), "utf8"))
+    .trim()
+    .split("\n")
+    .map((line) => JSON.parse(line).event);
+}
+
+describe("SDK migration log ordering", () => {
+  it("preserves invocation order for concurrent progress and completion writes", async () => {
+    const dataDir = await mkdtemp(
+      path.join(os.tmpdir(), "migration-log-order-"),
+    );
+    try {
+      const expected = Array.from(
+        { length: 100 },
+        (_, i) => `migration.progress.${i}`,
+      );
+      expected.push("migration.finished");
+      await Promise.all(
+        expected.map((event) =>
+          writeSdkMigrationLog(dataDir, {
+            writer: "desktop",
+            event,
+            outcome: "started",
+          }),
+        ),
+      );
+      expect(await events(dataDir)).toEqual(expected);
+    } finally {
+      await rm(dataDir, { recursive: true, force: true });
+    }
+  });
+
+  it("awaiting the terminal event also drains preceding progress writes", async () => {
+    const dataDir = await mkdtemp(
+      path.join(os.tmpdir(), "migration-log-drain-"),
+    );
+    const pending: Promise<void>[] = [];
+    try {
+      const expected = Array.from(
+        { length: 50 },
+        (_, i) => `migration.progress.${i}`,
+      );
+      for (const event of expected) {
+        pending.push(
+          writeSdkMigrationLog(dataDir, {
+            writer: "desktop",
+            event,
+            outcome: "started",
+          }),
+        );
+      }
+      await writeSdkMigrationLog(dataDir, {
+        writer: "desktop",
+        event: "migration.finished",
+        outcome: "succeeded",
+      });
+      expect(await events(dataDir)).toEqual([
+        ...expected,
+        "migration.finished",
+      ]);
+    } finally {
+      await Promise.all(pending);
+      await rm(dataDir, { recursive: true, force: true });
+    }
+  });
+
+  it("does not let a failed append block subsequent writes to the same home", async () => {
+    const dataDir = await mkdtemp(
+      path.join(os.tmpdir(), "migration-log-recovery-"),
+    );
+    try {
+      await writeFile(path.join(dataDir, "logs"), "not a directory");
+      await expect(
+        writeSdkMigrationLog(dataDir, {
+          writer: "desktop",
+          event: "migration.progress",
+          outcome: "started",
+        }),
+      ).resolves.toBeUndefined();
+      await rm(path.join(dataDir, "logs"));
+      await writeSdkMigrationLog(dataDir, {
+        writer: "desktop",
+        event: "migration.finished",
+        outcome: "succeeded",
+      });
+      expect(await events(dataDir)).toEqual(["migration.finished"]);
+    } finally {
+      await rm(dataDir, { recursive: true, force: true });
+    }
+  });
+});
diff --git a/memhub/packages/runtime-control/src/sdk-migration-log.ts b/memhub/packages/runtime-control/src/sdk-migration-log.ts
index 71bc831..ba4936e 100644
--- a/memhub/packages/runtime-control/src/sdk-migration-log.ts
+++ b/memhub/packages/runtime-control/src/sdk-migration-log.ts
@@ -33,6 +33,9 @@ export const SDK_MIGRATION_LOG_NAME = "sdk-migration.log";
 const MAX_BYTES = 4 * 1024 * 1024;
 const SAFE_TEXT = /^[A-Za-z0-9_.:/@+-]{0,160}$/;
 const MAX_ITEMS = 32;
+// Preserve invocation order within this process, including fire-and-forget
+// progress followed by an awaited terminal event. Homes remain independent.
+const pendingWrites = new Map<string, Promise<void>>();
 
 export function sdkMigrationLogPath(dataDir: string): string {
   return path.join(dataDir, "logs", SDK_MIGRATION_LOG_NAME);
@@ -103,17 +106,28 @@ export async function writeSdkMigrationLog(
       platform: process.platform,
       ...(entry.details === undefined ? {} : { details: entry.details }),
     });
-    const target = sdkMigrationLogPath(dataDir);
-    await mkdir(path.dirname(target), { recursive: true, mode: 0o700 });
-    const size = await stat(target).then(
-      (info) => info.size,
-      () => 0,
-    );
-    if (size > MAX_BYTES) await rename(target, `${target}.1`);
-    await appendFile(target, `${JSON.stringify(line)}\n`, {
-      encoding: "utf8",
-      mode: 0o600,
-    });
+    const target = path.resolve(sdkMigrationLogPath(dataDir));
+    const previous = pendingWrites.get(target) ?? Promise.resolve();
+    const pending = previous
+      .catch(() => undefined)
+      .then(async () => {
+        await mkdir(path.dirname(target), { recursive: true, mode: 0o700 });
+        const size = await stat(target).then(
+          (info) => info.size,
+          () => 0,
+        );
+        if (size > MAX_BYTES) await rename(target, `${target}.1`);
+        await appendFile(target, `${JSON.stringify(line)}\n`, {
+          encoding: "utf8",
+          mode: 0o600,
+        });
+      });
+    pendingWrites.set(target, pending);
+    try {
+      await pending;
+    } finally {
+      if (pendingWrites.get(target) === pending) pendingWrites.delete(target);
+    }
   } catch {
     // Never let the log decide whether a migration or a startup happens.
   }
diff --git a/memhub/scripts/generate_sbom.py b/memhub/scripts/generate_sbom.py
index 60212ba..c163ba0 100644
--- a/memhub/scripts/generate_sbom.py
+++ b/memhub/scripts/generate_sbom.py
@@ -36,7 +36,7 @@ MEMMY_LICENSE = (
 MEMMY_PROVENANCE = MEMMY_LICENSE.with_name("PROVENANCE.md")
 MEMMY_PIN = "25acd5f29fc7df73b943713abe61f1b2f9839887"
 HY_MEMORY_NAME = "hy-memory"
-HY_MEMORY_VERSION = "2.0.6"
+HY_MEMORY_VERSION = "2.0.8"
 
 
 class SbomError(RuntimeError):
diff --git a/memhub/scripts/install-docs.test.mjs b/memhub/scripts/install-docs.test.mjs
index 00147d8..24b5b6b 100644
--- a/memhub/scripts/install-docs.test.mjs
+++ b/memhub/scripts/install-docs.test.mjs
@@ -12,7 +12,7 @@ test("installation docs disclose publication and lifecycle boundaries", async ()
     "utf8",
   );
 
-  assert.match(readme, /`hy-memory==2\.0\.6`/);
+  assert.match(readme, /`hy-memory==2\.0\.8`/);
   assert.match(readme, /`v1\.0\.5` 的发行目标是 Linux x64\/arm64 Runtime/);
   assert.match(readme, /Linux 制品由 GitHub Actions/);
   assert.match(readme, /macOS.*Developer ID 签名、公证/s);
diff --git a/memhub/scripts/test_generate_sbom.py b/memhub/scripts/test_generate_sbom.py
index bda070d..0a11785 100644
--- a/memhub/scripts/test_generate_sbom.py
+++ b/memhub/scripts/test_generate_sbom.py
@@ -13,7 +13,7 @@ spec.loader.exec_module(module)
 
 
 def test_hy_memory_notice_comes_from_the_installed_dist_info(tmp_path: Path) -> None:
-    dist_info = tmp_path / "hy_memory-2.0.6.dist-info"
+    dist_info = tmp_path / "hy_memory-2.0.8.dist-info"
     licenses = dist_info / "licenses"
     licenses.mkdir(parents=True)
     (dist_info / "METADATA").write_text(
@@ -21,7 +21,7 @@ def test_hy_memory_notice_comes_from_the_installed_dist_info(tmp_path: Path) ->
             (
                 "Metadata-Version: 2.4",
                 "Name: hy-memory",
-                "Version: 2.0.6",
+                "Version: 2.0.8",
                 "License-Expression: Apache-2.0",
                 "",
             )
@@ -38,7 +38,7 @@ def test_hy_memory_notice_comes_from_the_installed_dist_info(tmp_path: Path) ->
     notice = module.hy_memory_notice(dist_info)
 
     assert notice["name"] == "hy-memory"
-    assert notice["version"] == "2.0.6"
+    assert notice["version"] == "2.0.8"
     assert notice["license"] == "Apache-2.0"
     assert notice["copyright"] == "Copyright (C) 2026 Tencent. All rights reserved."
     assert "Licensed under the Apache License" in notice["text"]
diff --git a/pyproject.toml b/pyproject.toml
index 83a4bf0..1588851 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "hy-memory"
-version = "2.0.6"
+version = "2.0.8"
 description = "HY Memory - Industrial-grade dual-system cognitive memory framework"
 readme = "README.md"
 requires-python = ">=3.9"
diff --git a/tests/coding/test_content_log_levels.py b/tests/coding/test_content_log_levels.py
new file mode 100644
index 0000000..6e8e7af
--- /dev/null
+++ b/tests/coding/test_content_log_levels.py
@@ -0,0 +1,140 @@
+# -*- coding: utf-8 -*-
+"""对话、记忆正文和模型输出只进 DEBUG，不进 INFO / WARNING。
+
+MemHub 等调用方默认以 INFO 运行；这些文本一旦进 INFO，就会写进常驻的
+hy_memory.log。这里逐条锁住曾经把正文打在 INFO / WARNING 的日志。
+"""
+
+import logging
+
+import pytest
+
+from hy_memory.agent.reconciler import ReconcileOp as AgentReconcileOp
+from hy_memory.agent.reconciler import _log_ops_detail
+from hy_memory.coding.extractor import CodingMemoryExtractor
+from hy_memory.coding.fact_reconciler import WorkingFactReconciler
+from hy_memory.coding.judge import classify_and_rewrite_queries
+from hy_memory.coding.reconciler import CodingMemoryReconciler
+from hy_memory.coding.types import CodingMemory, WorkingFact
+
+from tests.coding._helpers import MockLLMProvider
+
+SECRET = "SECRET-CONVERSATION-TEXT-42"
+
+
+def _above_debug(caplog):
+    return [
+        r.getMessage() for r in caplog.records
+        if r.name.startswith("hy_memory") and r.levelno > logging.DEBUG
+    ]
+
+
+def _debug(caplog):
+    return [
+        r.getMessage() for r in caplog.records
+        if r.name.startswith("hy_memory") and r.levelno == logging.DEBUG
+    ]
+
+
+def _assert_secret_only_at_debug(caplog):
+    assert not [m for m in _above_debug(caplog) if SECRET in m]
+    assert [m for m in _debug(caplog) if SECRET in m]
+
+
+@pytest.mark.asyncio
+class TestStoreInsertLogs:
+    async def test_experience_task_not_at_info(self, coding_store, caplog):
+        caplog.set_level(logging.DEBUG, logger="hy_memory")
+        await coding_store.insert(CodingMemory(
+            experience_id="e1", user_id="u1", task=SECRET, solution="S",
+        ))
+        info = [m for m in _above_debug(caplog) if "inserted experience_id=e1" in m]
+        assert info
+        _assert_secret_only_at_debug(caplog)
+
+    async def test_fact_statement_not_at_info(self, coding_store, caplog):
+        caplog.set_level(logging.DEBUG, logger="hy_memory")
+        await coding_store.insert_fact(WorkingFact(
+            fact_id="f1", user_id="u1", agent_id="a1", session_id="s1",
+            workspace_id="ws/a", statement=SECRET,
+        ))
+        info = [m for m in _above_debug(caplog) if "inserted fact_id=f1" in m]
+        assert info
+        _assert_secret_only_at_debug(caplog)
+
+
+class TestParseFailureWarnings:
+    """解析失败仍发 WARNING，但只带长度，模型输出进 DEBUG。"""
+
+    @pytest.mark.parametrize("parse", [
+        CodingMemoryExtractor._parse_output,
+        CodingMemoryReconciler._parse_ops,
+        WorkingFactReconciler._parse_ops,
+    ])
+    def test_warning_keeps_length_drops_text(self, parse, caplog):
+        caplog.set_level(logging.DEBUG, logger="hy_memory")
+        bad = f"not json {SECRET}"
+        parse(bad)
+        warnings = [
+            r.getMessage() for r in caplog.records
+            if r.levelno == logging.WARNING and "failed to parse" in r.getMessage()
+        ]
+        assert warnings and f"({len(bad)} chars)" in warnings[0]
+        _assert_secret_only_at_debug(caplog)
+
+
+@pytest.mark.asyncio
+class TestJudgeLogs:
+    async def test_rewrite_not_at_info(self, caplog):
+        caplog.set_level(logging.DEBUG, logger="hy_memory")
+        llm = MockLLMProvider(f"1\n{SECRET}")
+        result = await classify_and_rewrite_queries(["q"], llm)
+        assert result["rewrite_query"] == SECRET
+        _assert_secret_only_at_debug(caplog)
+
+    async def test_unrecognized_head_not_at_warning(self, caplog):
+        caplog.set_level(logging.DEBUG, logger="hy_memory")
+        llm = MockLLMProvider(f"{SECRET}\nrewrite")
+        result = await classify_and_rewrite_queries(["q"], llm)
+        assert result["ok"] is False
+        _assert_secret_only_at_debug(caplog)
+
+
+def test_reconciler_ops_detail_is_debug(caplog):
+    caplog.set_level(logging.DEBUG, logger="hy_memory")
+    _log_ops_detail([AgentReconcileOp(op="ADD", content=SECRET)])
+    _assert_secret_only_at_debug(caplog)
+
+
+class _StopAfterLogging(Exception):
+    pass
+
+
+def _add_client():
+    from types import SimpleNamespace
+    from unittest.mock import AsyncMock
+
+    from hy_memory.client import HyMemoryClient
+
+    client = HyMemoryClient.__new__(HyMemoryClient)
+    client._mode = "lite"
+    client._config = SimpleNamespace(coding=None)
+    writer = SimpleNamespace(
+        VERSION="test", write=AsyncMock(side_effect=_StopAfterLogging),
+    )
+    client._registry = SimpleNamespace(get_writer=AsyncMock(return_value=writer))
+    return client
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("data", [
+    SECRET,
+    [{"role": "user", "content": SECRET}],
+])
+async def test_add_does_not_log_input_at_info(data, caplog):
+    caplog.set_level(logging.DEBUG, logger="hy_memory")
+    client = _add_client()
+    with pytest.raises(_StopAfterLogging):
+        await client._async_add_impl(data, user_id="u1")
+    assert [m for m in _above_debug(caplog) if m.startswith("[add] writing")]
+    _assert_secret_only_at_debug(caplog)
diff --git a/tests/coding/test_memory_at.py b/tests/coding/test_memory_at.py
new file mode 100644
index 0000000..20731b7
--- /dev/null
+++ b/tests/coding/test_memory_at.py
@@ -0,0 +1,229 @@
+# -*- coding: utf-8 -*-
+"""经验与事实记下对话时间 memory_at，并以秒级 epoch 对外返回（与 chat 一致）。
+
+之前生产力链路丢了 add() 传进来的 memory_at，coding_memory_meta / working_fact
+也没有这一列，MemHub 里所有经验都显示"时间未知"。
+"""
+
+import sqlite3
+import time
+from datetime import datetime, timedelta, timezone
+from types import SimpleNamespace
+
+import pytest
+
+from hy_memory.client import HyMemoryClient
+from hy_memory.coding.extractor import CodingMemoryExtractor
+from hy_memory.coding.reconciler import CodingMemoryReconciler
+from hy_memory.coding.store import CodingMemoryStore, _CREATE_TABLE_SQL, _CREATE_FACT_TABLE_SQL
+from hy_memory.coding.types import CodingMemory, WorkingFact
+from hy_memory.coding.writer import CodingWriter
+
+from tests.coding._helpers import MockLLMProvider
+from tests.coding.test_list_and_stats_by_scene import _client, _unwrapped
+from tests.coding.test_write_scene_routing import _messages
+from tests.coding.test_writer_e2e import _coding_messages, _extract_resp, _reconcile_resp
+
+T1 = datetime(2026, 10, 1, 9, 30, tzinfo=timezone.utc)
+T2 = T1 + timedelta(days=3)
+
+
+def _epoch(value: datetime) -> int:
+    return int(value.timestamp())
+
+
+def _writer(store, embed, llm):
+    return CodingWriter(
+        store=store,
+        extractor=CodingMemoryExtractor(llm),
+        reconciler=CodingMemoryReconciler(store, embed, llm),
+        llm_provider=llm, embed_service=embed,
+    )
+
+
+@pytest.mark.asyncio
+class TestStore:
+    async def test_round_trip_as_epoch(self, coding_store):
+        await coding_store.insert(CodingMemory(
+            experience_id="e1", user_id="u", task="T", solution="S", memory_at=T1,
+        ))
+        await coding_store.insert_fact(WorkingFact(
+            fact_id="f1", user_id="u", statement="x", memory_at=T1,
+        ))
+        exp = await coding_store.get_by_id("e1", user_id="u")
+        fact = await coding_store.get_fact("f1", user_id="u")
+        assert _epoch(exp.memory_at) == _epoch(T1)
+        assert exp.to_dict()["memory_at"] == _epoch(T1)
+        assert fact.to_dict()["memory_at"] == _epoch(T1)
+
+    async def test_missing_time_stays_unknown(self, coding_store):
+        await coding_store.insert(CodingMemory(
+            experience_id="e1", user_id="u", task="T", solution="S",
+        ))
+        exp = await coding_store.get_by_id("e1", user_id="u")
+        assert exp.memory_at is None and exp.to_dict()["memory_at"] is None
+
+    @pytest.mark.parametrize("old, new", [(T1, T2), (T2, T1), (T2, None)])
+    async def test_update_keeps_the_later_time(self, coding_store, old, new):
+        await coding_store.insert(CodingMemory(
+            experience_id="e1", user_id="u", task="T", solution="S",
+            session_id="s", memory_at=old,
+        ))
+        await coding_store.update(CodingMemory(
+            experience_id="e1", user_id="u", task="T", solution="S2",
+            session_id="s", memory_at=new,
+        ))
+        exp = await coding_store.get_by_id("e1", user_id="u")
+        assert exp.solution == "S2"
+        assert _epoch(exp.memory_at) == _epoch(T2)
+
+    async def test_edit_keeps_the_time(self, coding_store):
+        await coding_store.insert(CodingMemory(
+            experience_id="e1", user_id="u", task="T", solution="S", memory_at=T1,
+        ))
+        await coding_store.edit_experience("e1", user_id="u", solution="S2")
+        exp = await coding_store.get_by_id("e1", user_id="u")
+        assert _epoch(exp.memory_at) == _epoch(T1)
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("variant", ["working", "coding"])
+async def test_existing_database_gains_the_column_on_open(coding_config, mock_embed, variant):
+    store = CodingMemoryStore(coding_config, embed_service=mock_embed, variant=variant)
+    # 先按老 schema（没有 memory_at）建库并放一行
+    conn = sqlite3.connect(store._db_path)
+    conn.executescript(_CREATE_TABLE_SQL.replace("    memory_at        INTEGER,\n", ""))
+    conn.executescript(_CREATE_FACT_TABLE_SQL.replace("    memory_at              INTEGER,\n", ""))
+    conn.execute(
+        "INSERT INTO coding_memory_meta (experience_id, user_id, task, solution, "
+        "created_at, updated_at) VALUES ('old', 'u', 'T', 'S', '2026-01-01T00:00:00', "
+        "'2026-01-01T00:00:00')"
+    )
+    conn.commit()
+    conn.close()
+
+    await store.initialize()
+    try:
+        old = await store.get_by_id("old", user_id="u")
+        assert old is not None and old.memory_at is None
+        await store.insert_fact(WorkingFact(
+            fact_id="f1", user_id="u", statement="x", memory_at=T1,
+        ))
+        assert _epoch((await store.get_fact("f1")).memory_at) == _epoch(T1)
+    finally:
+        await store.close()
+    # 再开一次不出错（ADD COLUMN 是幂等的）
+    again = CodingMemoryStore(coding_config, embed_service=mock_embed, variant=variant)
+    await again.initialize()
+    await again.close()
+
+
+@pytest.mark.asyncio
+class TestWriter:
+    def _llm(self):
+        return MockLLMProvider([
+            _extract_resp(
+                [{"task": "打包发布", "solution": "S", "confidence": 0.9}],
+                [{"statement": "发布脚本在 scripts/publish_package.sh"}],
+            ),
+            _reconcile_resp([{"draft_idx": 0, "action": "ADD"}]),
+        ])
+
+    async def test_conversation_time_reaches_both_layers(self, coding_store, mock_embed):
+        resp = await _writer(coding_store, mock_embed, self._llm()).write(
+            _coding_messages(), user_id="u", session_id="s", memory_at=T1,
+        )
+        exp = await coding_store.get_by_id(resp["experience_ids"][0], user_id="u")
+        fact = await coding_store.get_fact(resp["fact_ids"][0], user_id="u")
+        assert _epoch(exp.memory_at) == _epoch(T1)
+        assert _epoch(fact.memory_at) == _epoch(T1)
+
+    async def test_defaults_to_request_time_like_chat(self, coding_store, mock_embed):
+        before = int(time.time())
+        resp = await _writer(coding_store, mock_embed, self._llm()).write(
+            _coding_messages(), user_id="u", session_id="s",
+        )
+        exp = await coding_store.get_by_id(resp["experience_ids"][0], user_id="u")
+        fact = await coding_store.get_fact(resp["fact_ids"][0], user_id="u")
+        for stamp in (_epoch(exp.memory_at), _epoch(fact.memory_at)):
+            assert before <= stamp <= int(time.time()) + 1
+
+
+@pytest.mark.asyncio
+async def test_add_passes_the_time_to_the_productivity_writer(monkeypatch):
+    seen = {}
+
+    async def _judge(*args, **kwargs):
+        return "working"
+
+    class _Writer:
+        async def write(self, *args, **kwargs):
+            seen.update(kwargs)
+            return {"success": True, "experience_ids": [], "fact_ids": []}
+
+    async def _get_writer(*, variant):
+        return _Writer()
+
+    monkeypatch.setattr("hy_memory.client.classify_messages_scene", _judge)
+    ns = SimpleNamespace(
+        _config=SimpleNamespace(coding=SimpleNamespace(enable=True)),
+        _coding_llm_provider=object(),
+        _cache=None,
+        _get_coding_writer=_get_writer,
+        _normalize_scene=HyMemoryClient._normalize_scene,
+        _scene_variant=HyMemoryClient._scene_variant,
+        _parse_input=HyMemoryClient._parse_input,
+    )
+    await HyMemoryClient._async_add_impl(
+        ns, _messages(), user_id="u1", agent_id="a1", memory_at=T1,
+    )
+    assert seen["memory_at"] == T1
+
+
+@pytest.mark.asyncio
+async def test_list_get_and_search_return_epoch_seconds(coding_config, mock_embed):
+    coding = CodingMemoryStore(coding_config, embed_service=mock_embed, variant="coding")
+    await coding.initialize()
+    try:
+        await coding.insert(CodingMemory(
+            experience_id="e1", user_id="u", task="打包发布", solution="S", memory_at=T1,
+        ))
+        await coding.insert_fact(WorkingFact(
+            fact_id="f1", user_id="u", statement="打包发布", memory_at=T2,
+            derived_experience_ids=["e1"],
+        ))
+        client = _client([], {"coding": coding})
+        list_memories = _unwrapped("async_list_memories")
+
+        page = await list_memories(client, user_id="u", scene="coding")
+        assert page["experiences"][0]["memory_at"] == _epoch(T1)
+
+        merged = await list_memories(client, user_id="u", scene="all")
+        assert merged["memories"][0]["memory_at"] == _epoch(T1)
+
+        exp = await coding.get_by_id("e1", user_id="u")
+        got = HyMemoryClient._format_productivity_get(
+            exp, kind="experience", variant="coding", request_id="r", elapsed_ms=0,
+        )
+        assert got["memory_at"] == _epoch(T1)
+
+        client._embed_service = mock_embed
+        client._coding_store = coding
+
+        async def _ready(variant="working"):
+            return None
+
+        client._get_coding_writer = _ready
+        client._cache = None
+        resp = await HyMemoryClient._search_coding(
+            client, target_query="打包发布", user_ids=["u"], workspace_id=None,
+            branch=None, limit=5, working_fact_min_score=0.0,
+            working_experience_min_score=0.0, request_id="r",
+            t0=time.perf_counter(), variant="coding",
+        )
+        assert resp["facts"][0]["memory_at"] == _epoch(T2)
+        assert resp["experiences"][0]["memory_at"] == _epoch(T1)
+        unified = HyMemoryClient._search_lane_items("coding", resp)
+        assert {item["memory_at"] for item in unified} == {_epoch(T1), _epoch(T2)}
+    finally:
+        await coding.close()
diff --git a/tests/test_async_executor_context.py b/tests/test_async_executor_context.py
new file mode 100644
index 0000000..5fb6a47
--- /dev/null
+++ b/tests/test_async_executor_context.py
@@ -0,0 +1,51 @@
+"""The guarded to_thread adapter preserves asyncio.to_thread context semantics."""
+import asyncio
+import contextvars
+import importlib.util
+from pathlib import Path
+import unittest
+
+
+# This helper is dependency-free; exercise it without initializing the SDK.
+_spec = importlib.util.spec_from_file_location(
+    "async_executor_under_test",
+    Path(__file__).resolve().parents[1] / "hy_memory/utils/async_executor.py",
+)
+_executor = importlib.util.module_from_spec(_spec)
+_spec.loader.exec_module(_executor)
+
+
+class ThreadContextTests(unittest.IsolatedAsyncioTestCase):
+    async def test_copies_context_and_does_not_propagate_worker_changes(self):
+        request_id = contextvars.ContextVar("request_id", default="missing")
+        request_id.set("caller")
+
+        def work(prefix, *, suffix):
+            value = request_id.get()
+            request_id.set("worker")
+            return prefix + value + suffix
+
+        result = await _executor.to_thread_guarded(work, "[", suffix="]")
+        self.assertEqual(result, "[caller]")
+        self.assertEqual(request_id.get(), "caller")
+
+    async def test_concurrent_callers_keep_separate_contexts(self):
+        request_id = contextvars.ContextVar("request_id", default="missing")
+
+        async def call(value):
+            request_id.set(value)
+            await asyncio.sleep(0)
+            return await _executor.to_thread_guarded(request_id.get)
+
+        self.assertEqual(await asyncio.gather(call("a"), call("b")), ["a", "b"])
+
+    async def test_context_is_copied_when_coroutine_runs(self):
+        request_id = contextvars.ContextVar("request_id", default="missing")
+        request_id.set("created")
+        pending = _executor.to_thread_guarded(request_id.get)
+        request_id.set("awaited")
+        self.assertEqual(await pending, "awaited")
+
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/tests/test_embed_warning_log.py b/tests/test_embed_warning_log.py
index bb561e5..ae4e90e 100644
--- a/tests/test_embed_warning_log.py
+++ b/tests/test_embed_warning_log.py
@@ -90,7 +90,10 @@ class TestEmbedInfoLog:
             assert payload["provider"] == "openai"
             assert payload["model"] == "text-embedding-test"
             assert payload["api"] == "https://example.com/v1/embeddings"
-            assert payload["request"]["input"] == ["hello world"]
+            # 待向量化的原文不进 INFO，只记每条字符数
+            assert "input" not in payload["request"]
+            assert payload["request"]["input_chars"] == [11]
+            assert "hello world" not in info_msgs[0]
             assert payload["request"]["dimensions"] == 8
             assert payload["response"]["vector_dims"] == [8]
             assert payload["response"]["vector_count"] == 1
diff --git a/tests/test_literal_special_token_counting.py b/tests/test_literal_special_token_counting.py
new file mode 100644
index 0000000..a2eb461
--- /dev/null
+++ b/tests/test_literal_special_token_counting.py
@@ -0,0 +1,29 @@
+"""User text may quote tokenizer control spellings without using control tokens."""
+
+import pytest
+import tiktoken
+
+from hy_memory.core.embed_service import _count_tokens, _truncate_for_embedding
+from hy_memory.utils.token_count import count_tokens
+
+
+@pytest.mark.parametrize('counter', [count_tokens, _count_tokens])
+@pytest.mark.parametrize('marker', ['<|endoftext|>', '<|fim_prefix|>'])
+def test_literal_special_tokens_are_counted_as_ordinary_text(marker, counter):
+    text = marker + '你好世界' * 20
+    expected = len(tiktoken.get_encoding('cl100k_base').encode_ordinary(text))
+    assert counter(text) == expected
+
+
+def test_literal_special_token_does_not_bypass_embedding_limit():
+    text = '<|endoftext|>' + '你好世界' * 3000
+    encoder = tiktoken.get_encoding('cl100k_base')
+    truncated = _truncate_for_embedding(text)
+    assert truncated != text
+    assert len(encoder.encode_ordinary(truncated)) <= 8192 - 64
+    assert truncated.startswith('<|endoftext|>')
+
+
+def test_short_literal_special_token_text_is_preserved():
+    text = 'Explain <|endoftext|> as a literal string.'
+    assert _truncate_for_embedding(text) == text
diff --git a/tests/test_pipeline_observability.py b/tests/test_pipeline_observability.py
index 40f5b30..e81f81a 100644
--- a/tests/test_pipeline_observability.py
+++ b/tests/test_pipeline_observability.py
@@ -1,4 +1,4 @@
-"""Pipeline log（始终）与 trace（可关）分离。"""
+"""Pipeline log（默认开，可关）与 trace（可关）分离。"""
 
 import asyncio
 import json
@@ -10,6 +10,7 @@ from hy_memory.config import MemoryConfig
 from hy_memory.data.cache_sqlite import SqliteCache
 from hy_memory.utils.pipeline_observability import (
     is_llm_content_logging_enabled,
+    is_pipeline_log_enabled,
     is_pipeline_trace_enabled,
     resolve_pipeline_log_dir,
 )
@@ -39,6 +40,14 @@ class TestPipelineObservabilityConfig:
         monkeypatch.setenv("MEMORY_DATA_DIR", str(tmp_path / "data"))
         assert resolve_pipeline_log_dir() == str(tmp_path / "data" / "logs" / "pipeline")
 
+    def test_log_enabled_default(self, monkeypatch):
+        monkeypatch.delenv("MEMORY_PIPELINE_LOG_ENABLED", raising=False)
+        assert is_pipeline_log_enabled() is True
+
+    def test_log_disabled(self, monkeypatch):
+        monkeypatch.setenv("MEMORY_PIPELINE_LOG_ENABLED", "false")
+        assert is_pipeline_log_enabled() is False
+
     def test_trace_enabled_default(self, monkeypatch):
         monkeypatch.delenv("MEMORY_PIPELINE_TRACE_ENABLED", raising=False)
         assert is_pipeline_trace_enabled() is True
@@ -91,6 +100,98 @@ class TestObservabilityHookBehavior:
 
         run(_test())
 
+    def test_log_flag_on_writes_jsonl(self, env_setup, monkeypatch):
+        """显式开启与默认一致：照常写 JSONL。"""
+        monkeypatch.setenv("MEMORY_PIPELINE_LOG_ENABLED", "true")
+        monkeypatch.setenv("MEMORY_PIPELINE_TRACE_ENABLED", "false")
+
+        async def _test():
+            from datetime import date
+            from hy_memory.utils.pipeline_observability import (
+                drain_pipeline_log_tasks,
+                install_pipeline_observability,
+            )
+
+            config = MemoryConfig.from_dict({"cache": {"backend": "sqlite"}})
+            cache = SqliteCache(config)
+            await cache.initialize()
+            install_pipeline_observability(cache, log_label="test")
+            await cache.store_pipeline_log(
+                request_id="req-on", user_id="u1", agent_id="a1", step="S1_EMBED",
+            )
+            await drain_pipeline_log_tasks()
+
+            log_file = Path(env_setup) / "default" / f"{date.today().isoformat()}.log"
+            assert json.loads(log_file.read_text(encoding="utf-8").strip())["step"] == "S1_EMBED"
+            await cache.close()
+
+        run(_test())
+
+    def test_log_off_trace_off_is_noop(self, env_setup, monkeypatch):
+        """两者都关：不包装、不写文件、不写库，store_pipeline_log 直接返回 True。"""
+        monkeypatch.setenv("MEMORY_PIPELINE_LOG_ENABLED", "false")
+        monkeypatch.setenv("MEMORY_PIPELINE_TRACE_ENABLED", "false")
+
+        async def _test():
+            from hy_memory.utils.pipeline_observability import (
+                _noop_store_pipeline_log,
+                drain_pipeline_log_tasks,
+                install_pipeline_observability,
+            )
+
+            config = MemoryConfig.from_dict({"cache": {"backend": "sqlite"}})
+            cache = SqliteCache(config)
+            await cache.initialize()
+            original_store = cache.store_pipeline_log
+            install_pipeline_observability(cache, log_label="test")
+            assert cache.store_pipeline_log is _noop_store_pipeline_log
+
+            ok = await cache.store_pipeline_log(
+                request_id="req-off", user_id="ns__u1", agent_id="a1",
+                step="S1_EMBED", prompt="p", response="r",
+            )
+            await drain_pipeline_log_tasks()
+
+            assert ok is True
+            assert not Path(env_setup).exists()
+            # 原始 store 仍能写库，说明表本身可用；空操作确实没写
+            assert await cache.get_pipeline_logs(request_id="req-off") == []
+            await original_store(
+                request_id="req-probe", user_id="u1", agent_id="a1",
+                step="PROBE", prompt="", response="",
+            )
+            assert len(await cache.get_pipeline_logs(request_id="req-probe")) == 1
+            await cache.close()
+
+        run(_test())
+
+    def test_log_off_trace_on_writes_db_only(self, env_setup, monkeypatch):
+        """只关 Log：不写 JSONL，Trace 照常落库。"""
+        monkeypatch.setenv("MEMORY_PIPELINE_LOG_ENABLED", "false")
+        monkeypatch.setenv("MEMORY_PIPELINE_TRACE_ENABLED", "true")
+
+        async def _test():
+            from hy_memory.utils.pipeline_observability import (
+                drain_pipeline_log_tasks,
+                install_pipeline_observability,
+            )
+
+            config = MemoryConfig.from_dict({"cache": {"backend": "sqlite"}})
+            cache = SqliteCache(config)
+            await cache.initialize()
+            install_pipeline_observability(cache, log_label="test")
+            await cache.store_pipeline_log(
+                request_id="req-db", user_id="u1", agent_id="a1", step="S1_EMBED",
+            )
+            await drain_pipeline_log_tasks()
+
+            assert not Path(env_setup).exists()
+            logs = await cache.get_pipeline_logs(request_id="req-db")
+            assert [row["step"] for row in logs] == ["S1_EMBED"]
+            await cache.close()
+
+        run(_test())
+
     def test_store_returns_before_slow_persist(self, env_setup, monkeypatch):
         """慢 DB / 慢磁盘不得堵住 await store_pipeline_log。"""
         monkeypatch.setenv("MEMORY_PIPELINE_TRACE_ENABLED", "true")