chore(release): bump MemHub build number to 18

本包代码与 build 17 的 commit 不同但字节不同，沿用 17 会触发
local-install 的 installed_version_conflict（同 build_number 不同字节拒装）。
升到 18，保证能从 build 17 平滑升级。

hy_memory/client.py

@@ -495,8 +495,11 @@ class HyMemoryClient:
 
         # 进程级初始化日志强制不带业务 request_id（哨兵 "-"），避免污染 App
         # 请求上下文里按 request_id 的过滤。
+        #
+        # 用 debug 而非 info：全组件冷启动本就该是进程级一次性事件，
+        # 一旦有人在高频路径上重建 client，info 会把这条噪音放大几十万倍。
         with request_id_scope():
-            logger.info(
+            logger.debug(
                 f"Initializing HyMemoryClient: mode={self._mode} "
                 f"enable_graph={self.enable_graph}"
             )
@@ -523,7 +526,7 @@ class HyMemoryClient:
             # 同步初始化所有组件
             self._loop_thread.run(self._initialize())
 
-            logger.info("HyMemoryClient ready")
+            logger.debug("HyMemoryClient ready")
 
     async def _initialize(self):
         """初始化所有核心组件"""
@@ -542,7 +545,7 @@ class HyMemoryClient:
         loop.set_default_executor(
             concurrent.futures.ThreadPoolExecutor(max_workers=_max_workers)
         )
-        logger.info(f"[init] Default ThreadPoolExecutor max_workers={_max_workers}")
+        logger.debug(f"[init] Default ThreadPoolExecutor max_workers={_max_workers}")
 
         # 1. EmbedService（注入发生在 ComponentFactory / pipeline 预热之前）
         self._embed_service = (
@@ -557,7 +560,7 @@ class HyMemoryClient:
         # 2. VectorStore（根据配置选择后端）
         self._vector_store = create_vector_store(self._config)
         await self._vector_store.initialize()
-        logger.info(
+        logger.debug(
             f"[init] VectorStore ready: provider={self._config.vector_store.provider} "
             f"collection={self._vector_store._collection_name}"
         )
@@ -574,7 +577,7 @@ class HyMemoryClient:
                     f"GraphStore ({self._config.graph_store.provider}) init failed. "
                     f"Use mode='pro' instead of mode='ultra' to disable Graph, or check logs for details."
                 )
-            logger.info(
+            logger.debug(
                 f"[init] GraphStore ready: provider={self._config.graph_store.provider}"
             )
 
@@ -594,10 +597,10 @@ class HyMemoryClient:
         if self._owns_runtime:
             self._cache = create_cache(self._config)
             await self._cache.initialize()
-            logger.info(f"[init] Cache ready (backend={self._config.cache.backend})")
+            logger.debug(f"[init] Cache ready (backend={self._config.cache.backend})")
         else:
             self._cache = self._runtime.cache
-            logger.info("[init] Cache reused from SharedRuntime")
+            logger.debug("[init] Cache reused from SharedRuntime")
 
         self._registry._shared_cache = self._cache
 
@@ -610,14 +613,14 @@ class HyMemoryClient:
         # 6. 预热组件（确保首次 add/search 不需要等待初始化）
         await self._registry.get_writer()
         await self._registry.get_reader()
-        logger.info(f"[init] Writer & Reader ready (mode={self._mode})")
+        logger.debug(f"[init] Writer & Reader ready (mode={self._mode})")
 
         await self._initialize_local_reranker()
 
         if self._mode == "ultra":
             try:
                 await self._registry.get_system2_writer()
-                logger.info("[init] System2Writer ready (ultra mode)")
+                logger.debug("[init] System2Writer ready (ultra mode)")
             except Exception as e:
                 logger.warning(f"[init] System2Writer warmup failed: {e}")
 
@@ -626,7 +629,7 @@ class HyMemoryClient:
         if self._config.history.enable:
             self._history_store = HistoryStore(self._config)
             await self._history_store.initialize()
-            logger.info("[init] HistoryStore ready")
+            logger.debug("[init] HistoryStore ready")
 
         # 8. MetricsCollector
         # - solo 模式：bind cache 到 metrics 单例 + 启 background flush
@@ -636,9 +639,9 @@ class HyMemoryClient:
             _metrics = MetricsCollector.get()
             _metrics.bind_cache(self._cache)
             await _metrics.start_background_tasks()
-            logger.info("[init] MetricsCollector ready")
+            logger.debug("[init] MetricsCollector ready")
         else:
-            logger.info("[init] MetricsCollector already bound by SharedRuntime")
+            logger.debug("[init] MetricsCollector already bound by SharedRuntime")
 
         # 9. Coding memory 链路占位（lazy 初始化）
         #    仅在 add(messages_with_tool) 首次命中 is_coding=True 时才真正初始化。
@@ -691,7 +694,7 @@ class HyMemoryClient:
                 "Local rerank is enabled but initialization failed: "
                 f"model={reranker.model_name!r}, error={exc}"
             ) from exc
-        logger.info(
+        logger.debug(
             "[init] Local reranker ready: model=%s device=%s "
             "load_ms=%s warmup_ms=%s",
             reranker.model_name,

hy_memory/snapshot.py

@@ -15,6 +15,15 @@ class DataPlaneSnapshotRejected(RuntimeError):
         self.code = code
 
 
+# Runtime logs live beside the data plane but are not data: they are written by
+# a rotating handler, carry no recovery value, and can grow to gigabytes. The
+# migration snapshot is a recovery artifact, so copying them both doubles the
+# disk cost and makes every migration run fail on a nearly-full volume.
+# Only top-level entries are excluded so a business directory that happens to be
+# named "logs" deeper in the tree is still copied.
+SNAPSHOT_EXCLUDED_TOP_LEVEL: frozenset[str] = frozenset({"logs"})
+
+
 @dataclass(frozen=True)
 class DataPlaneSnapshot:
     source_root: Path
@@ -64,6 +73,14 @@ def snapshot_local_data_plane(
         for root, directories, files in os.walk(source, followlinks=False):
             root_path = Path(root)
             relative = root_path.relative_to(source)
+            if relative == Path("."):
+                # Prune before descending: an excluded directory is never walked,
+                # so its files are neither checked nor copied.
+                directories[:] = [
+                    name
+                    for name in directories
+                    if name not in SNAPSHOT_EXCLUDED_TOP_LEVEL
+                ]
             target_directory = target / relative
             target_directory.mkdir(parents=True, exist_ok=True)
             for name in directories:

hy_memory/utils/log_setup.py

@@ -3,8 +3,8 @@
 HY Memory - 日志系统配置
 
 标准 Python logging + TimedRotatingFileHandler，
-每天一个文件，保留 15 天。支持 request_id 注入，
-可按 request_id grep 日志过滤完整请求链路。
+每天一个文件，默认保留 3 天（可用 MEMORY_LOG_RETENTION_DAYS 覆盖）。
+支持 request_id 注入，可按 request_id grep 日志过滤完整请求链路。
 
 Usage:
     from hy_memory.utils.log_setup import setup_logging, set_request_id
@@ -18,6 +18,7 @@ Usage:
 """
 
 import contextvars
+import datetime as _datetime
 import logging
 import os
 import sys
@@ -25,6 +26,113 @@ from logging.handlers import TimedRotatingFileHandler
 from pathlib import Path
 from typing import Optional
 
+# ================================================================
+# 日志保留策略
+# ================================================================
+
+# 默认只保留 3 天。INFO 级日志单日可涨到 2 GB 以上，15 天窗口在真实使用中
+# 会累积到 10 GB 量级，而这份体积会被迁移快照整份复制一遍。
+DEFAULT_LOG_RETENTION_DAYS = 3
+LOG_RETENTION_ENV = "MEMORY_LOG_RETENTION_DAYS"
+_ROTATION_SUFFIX_FORMAT = "%Y-%m-%d"
+
+
+def _is_rotation_date(value: str) -> bool:
+    """Return True when value is a rotation date suffix such as 2026-10-19."""
+    if len(value) != len("2026-10-19"):
+        return False
+    try:
+        _datetime.datetime.strptime(value, _ROTATION_SUFFIX_FORMAT)
+    except ValueError:
+        return False
+    return True
+
+
+def resolve_retention_days(explicit: Optional[int] = None) -> int:
+    """Resolve the retention window: explicit arg > env var > default."""
+    if explicit is not None:
+        try:
+            return max(1, int(explicit))
+        except (TypeError, ValueError):
+            pass
+    raw = os.getenv(LOG_RETENTION_ENV, "").strip()
+    if raw:
+        try:
+            return max(1, int(raw))
+        except ValueError:
+            pass
+    return DEFAULT_LOG_RETENTION_DAYS
+
+
+class GzipAwareTimedRotatingFileHandler(TimedRotatingFileHandler):
+    """TimedRotatingFileHandler that still expires files after they are gzipped.
+
+    The stdlib implementation builds its deletion list by matching the active
+    file name against a date-suffixed pattern. Once an operator gzips a rotated
+    file the suffix no longer matches, so that file drops out of the list; when
+    fewer files than ``backupCount`` match, ``getFilesToDelete()`` returns ``[]``
+    and nothing is ever deleted again. The directory then grows forever while
+    looking correctly configured.
+    """
+
+    def getFilesToDelete(self) -> list[str]:  # noqa: N802 - stdlib naming
+        directory, base = os.path.split(self.baseFilename)
+        if not os.path.isdir(directory):
+            return []
+        prefix = base + "."
+        candidates: list[str] = []
+        for name in os.listdir(directory):
+            if not name.startswith(prefix):
+                continue
+            suffix = name[len(prefix) :]
+            if suffix.endswith(".gz"):
+                suffix = suffix[: -len(".gz")]
+            if _is_rotation_date(suffix):
+                candidates.append(os.path.join(directory, name))
+        candidates.sort()
+        if self.backupCount <= 0 or len(candidates) < self.backupCount:
+            return []
+        return candidates[: len(candidates) - self.backupCount]
+
+
+def purge_expired_logs(
+    log_dir: str,
+    retention_days: int,
+    *,
+    base_name: str = "hy_memory.log",
+    today: Optional[_datetime.date] = None,
+) -> list[str]:
+    """Delete rotated log files older than the retention window.
+
+    Runs once at startup in addition to the handler's own rotation cleanup: a
+    directory that already accumulated weeks of files (or gzipped files the
+    handler cannot match) would otherwise never converge. Age comes from the
+    date in the file name, not mtime, because copied files carry a new mtime.
+    """
+    directory = Path(log_dir)
+    if not directory.is_dir():
+        return []
+    reference = today or _datetime.date.today()
+    prefix = base_name + "."
+    removed: list[str] = []
+    for path in directory.iterdir():
+        if not path.name.startswith(prefix) or not path.is_file():
+            continue
+        suffix = path.name[len(prefix) :]
+        if suffix.endswith(".gz"):
+            suffix = suffix[: -len(".gz")]
+        if not _is_rotation_date(suffix):
+            continue
+        written = _datetime.datetime.strptime(suffix, _ROTATION_SUFFIX_FORMAT).date()
+        if (reference - written).days > retention_days:
+            try:
+                path.unlink()
+            except OSError:
+                continue
+            removed.append(str(path))
+    return removed
+
+
 # ================================================================
 # request_id 上下文变量 (协程安全)
 # ================================================================
@@ -184,7 +292,7 @@ def _default_log_dir() -> str:
 
 def setup_logging(
     log_dir: Optional[str] = None,
-    retention_days: int = 15,
+    retention_days: Optional[int] = None,
     level: Optional[int] = None,
 ) -> None:
     """
@@ -198,7 +306,8 @@ def setup_logging(
 
     Args:
         log_dir:        日志目录。默认 {MEMORY_DATA_DIR}/logs/
-        retention_days: 日志保留天数，默认 15
+        retention_days: 日志保留天数。默认读 MEMORY_LOG_RETENTION_DAYS，
+                        再缺省为 DEFAULT_LOG_RETENTION_DAYS（3 天）
         level:          日志级别。默认从环境变量 MEMORY_LOG_LEVEL 读取，缺省 INFO
 
     可多次调用，但只有首次生效。
@@ -208,6 +317,8 @@ def setup_logging(
         return
     _logging_configured = True
 
+    resolved_retention = resolve_retention_days(retention_days)
+
     if level is None:
         level = getattr(logging, os.getenv("MEMORY_LOG_LEVEL", "INFO").upper(), logging.INFO)
 
@@ -215,6 +326,9 @@ def setup_logging(
     resolved_dir = log_dir or _default_log_dir()
     os.makedirs(resolved_dir, exist_ok=True)
 
+    # 先按策略清一次历史文件：轮转清理只在滚动时触发，已经积压的目录不会自己收敛
+    purge_expired_logs(resolved_dir, resolved_retention)
+
     # hy_memory 根 logger
     root_logger = logging.getLogger("hy_memory")
     root_logger.setLevel(level)
@@ -231,11 +345,11 @@ def setup_logging(
 
     # ---- 文件 Handler ----
     log_file = os.path.join(resolved_dir, "hy_memory.log")
-    file_handler = TimedRotatingFileHandler(
+    file_handler = GzipAwareTimedRotatingFileHandler(
         filename=log_file,
         when="midnight",
         interval=1,
-        backupCount=retention_days,
+        backupCount=resolved_retention,
         encoding="utf-8",
     )
     # 滚动后文件名: hy_memory.log.2026-03-21

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

@@ -9,6 +9,7 @@ from __future__ import annotations
 
 import asyncio
 import hashlib
+import logging
 import os
 import threading
 from collections.abc import Callable, Coroutine
@@ -40,6 +41,8 @@ BUILTIN_RETRIEVAL_PROFILE_ID = "builtin-retrieval"
 _LOCAL_ONNX_PLACEHOLDER_URL = "http://127.0.0.1/local-onnx"
 DEFAULT_RELEVANCE_FLOOR = 0.35
 
+logger = logging.getLogger("one_memory.hy_sdk_runtime")
+
 
 @dataclass
 class _RuntimeEntry:
@@ -120,6 +123,15 @@ class ClientLease:
             if self._closed:
                 return
             self._closed = True
+            if _lease_trace_enabled():
+                # 一个进程只应持有极少数 client lease。若这里被高频调用，
+                # 说明某条请求路径在请求结束时释放了共享句柄，下一次调用
+                # 只能整份重建 client（连带 3 套组件与线程池）。
+                logger.warning(
+                    "sdk client lease released: data_dir=%s",
+                    self.data_dir,
+                    stack_info=True,
+                )
             try:
                 self.client.close()
             finally:
@@ -132,6 +144,15 @@ class ClientLease:
         self.close()
 
 
+def _lease_trace_enabled() -> bool:
+    return os.getenv("ONE_MEMORY_TRACE_CLIENT_LEASE", "").strip() not in {
+        "",
+        "0",
+        "false",
+        "False",
+    }
+
+
 class SdkClientHandle:
     """Lazy async owner for one profile-bound public SDK client lease."""
 

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

@@ -9,7 +9,6 @@ import sqlite3
 from collections.abc import Callable
 from contextlib import suppress
 from dataclasses import dataclass
-from datetime import UTC, datetime, timedelta
 from pathlib import Path
 
 from one_memory.database_access import main_database
@@ -28,11 +27,11 @@ from one_memory.sdk_data_plane_migration import (
 )
 from one_memory.sdk_migration_bundle import inspect_sdk_migration_bundle
 from one_memory.sdk_migration_inventory import (
+    LegacyRawInventory,
     SdkMigrationClient,
     execute_memory_migration_plan,
     execute_raw_migration_plan,
     read_legacy_memory_inventory,
-    read_legacy_raw_inventory,
     read_sdk_memory_inventory,
     read_sdk_raw_inventory,
     record_memory_migration_plan,
@@ -414,20 +413,13 @@ async def run_sdk_data_plane_backfill(
         legacy=legacy_memories,
         sdk=sdk_memories,
     )
-    with main_database(database_path).read(row_factory=sqlite3.Row) as database:
-        started = database.execute(
-            "SELECT created_at FROM sdk_data_plane_migrations WHERE migration_id=?",
-            (migration_id,),
-        ).fetchone()
-    if started is None:
-        raise DataPlaneMigrationError("sdk_data_plane_migration_start_missing")
-    started_at = datetime.fromisoformat(str(started["created_at"]))
-    if started_at.tzinfo is None:
-        started_at = started_at.replace(tzinfo=UTC)
-    legacy_raw = read_legacy_raw_inventory(
-        database_path,
-        cutoff=started_at.astimezone(UTC) - timedelta(days=3),
-    )
+    # raw 全丢：迁移只搬记忆，不再搬原始对话。
+    #
+    # 仍然走同一套 plan/execute 路径，classify_raw_inventory 的判定规则不动：
+    # 新数据面里已有的 raw 因为拿不到 legacy 对应项，会按原有规则被判为窗口外
+    # 并删除。这里只做一件事——不再从 legacy 读取、不再把三天窗口内的对话
+    # 重建成一份新的 raw，省掉几十万行的读写与重建成本。
+    legacy_raw: tuple[LegacyRawInventory, ...] = ()
     sdk_raw = await read_sdk_raw_inventory(client)
     raw_decisions = record_raw_migration_plan(
         store,

memhub/apps/core/tests/test_sdk_data_plane_migration.py

@@ -1028,28 +1028,16 @@ def test_offline_backfill_reaches_ready_without_committing_cutover(store):
 
     assert result.state == "ready"
     assert result.phase == "reconciled"
-    assert result.counts == {"verified": 2}
+    assert result.counts == {"verified": 1}
     assert [item["memory_id"] for item in client.imported] == ["memory-1"]
-    assert len(client.imported_raw) == 1
-    assert client.imported_raw[0]["content"] == "[user]: verbatim\n[assistant]: reply"
-    migrated_raw_id = client.imported_raw[0]["raw_id"]
+    # raw 全丢：迁移只搬记忆，不再把 legacy 的对话重建成新数据面的 raw
+    assert client.imported_raw == []
     assert store.authority() is MemoryAuthority.LEGACY
-    # Simulate a process dying after both SDK side effects but before startup
-    # could finish. The next inventory observes them as matched rather than
+    # Simulate a process dying after the SDK side effect but before startup
+    # could finish. The next inventory observes it as matched rather than
     # import_exact; that reclassification must resume the settled ledger
     # instead of conflicting with its original operation key.
     client.memories = (_record("memory-1", content="preference"),)
-    client.raw = (
-        HySdkRawRecord(
-            raw_id=migrated_raw_id,
-            content="[user]: verbatim\n[assistant]: reply",
-            user_id="local",
-            agent_id="all-apps",
-            session_id="session",
-            source="cursor",
-            custom={"input_kind": "legacy_turn"},
-        ),
-    )
     resumed = asyncio.run(
         run_sdk_data_plane_backfill(
             database_path=database,
@@ -1064,9 +1052,9 @@ def test_offline_backfill_reaches_ready_without_committing_cutover(store):
         )
     )
     assert resumed.state == "ready"
-    assert resumed.counts == {"verified": 2}
+    assert resumed.counts == {"verified": 1}
     assert len(client.imported) == 1
-    assert len(client.imported_raw) == 1
+    assert client.imported_raw == []
 
 
 def test_offline_backfill_blocks_cutover_when_an_item_is_quarantined(store):

memhub/package.json

@@ -2,7 +2,7 @@
   "name": "one-memory",
   "version": "1.0.5",
   "release": {
-    "build_number": 17,
+    "build_number": 18,
     "macos_min_version": "13.0",
     "app_update": "manual",
     "runtime_manifest_schema": 1,

tests/test_data_plane_snapshot.py

@@ -39,6 +39,37 @@ def test_snapshot_copies_and_verifies_local_sdk_tree(tmp_path):
     assert result.total_bytes == len(b"vector") + len(b"cache")
 
 
+def test_snapshot_skips_top_level_runtime_logs(tmp_path):
+    """Runtime logs sit beside the data plane but are not data.
+
+    Copying them doubled a 15.8 GB migration snapshot, most of it rotated log
+    files, and made migrations fail with ENOSPC on an otherwise healthy volume.
+    Only the top-level directory is pruned: a business directory that happens to
+    be called "logs" deeper in the tree must still be copied.
+    """
+    source = tmp_path / "source"
+    (source / "logs").mkdir(parents=True)
+    (source / "logs" / "hy_memory.log").write_bytes(b"x" * 4096)
+    (source / "logs" / "hy_memory.log.2026-10-19").write_bytes(b"y" * 4096)
+    (source / "vector" / "logs").mkdir(parents=True)
+    (source / "vector" / "logs" / "keep.bin").write_bytes(b"keep")
+    (source / "cache.sqlite3").write_bytes(b"cache")
+    target = tmp_path / "snapshot"
+
+    result = snapshot_local_data_plane(
+        source_root=source,
+        target_root=target,
+        writers_stopped=True,
+    )
+
+    assert {item["path"] for item in result.files} == {
+        "cache.sqlite3",
+        "vector/logs/keep.bin",
+    }
+    assert not (target / "logs").exists()
+    assert result.total_bytes == len(b"keep") + len(b"cache")
+
+
 def test_snapshot_refuses_active_writers(tmp_path):
     source = tmp_path / "source"
     source.mkdir()

tests/test_log_setup.py

@@ -222,3 +222,72 @@ class TestSingleLineAndJsonFormat:
         monkeypatch.delenv("MEMORY_LOG_FORMAT", raising=False)
         fmt = log_setup._make_formatter()
         assert isinstance(fmt, log_setup.SingleLineFormatter)
+
+
+class TestLogRetention:
+    """保留窗口必须能真正收敛，而不是只在配置里写着 3 天。"""
+
+    def test_default_retention_is_three_days(self, monkeypatch):
+        monkeypatch.delenv("MEMORY_LOG_RETENTION_DAYS", raising=False)
+        assert log_setup.resolve_retention_days() == 3
+
+    def test_env_overrides_default(self, monkeypatch):
+        monkeypatch.setenv("MEMORY_LOG_RETENTION_DAYS", "7")
+        assert log_setup.resolve_retention_days() == 7
+
+    def test_explicit_argument_wins_over_env(self, monkeypatch):
+        monkeypatch.setenv("MEMORY_LOG_RETENTION_DAYS", "7")
+        assert log_setup.resolve_retention_days(2) == 2
+
+    def test_invalid_env_falls_back_to_default(self, monkeypatch):
+        monkeypatch.setenv("MEMORY_LOG_RETENTION_DAYS", "not-a-number")
+        assert log_setup.resolve_retention_days() == 3
+
+    def test_purge_removes_expired_including_gzipped_files(self, tmp_path):
+        import datetime as _dt
+        from pathlib import Path
+
+        log_dir = tmp_path / "logs"
+        log_dir.mkdir()
+        expired = ["hy_memory.log.2026-10-01", "hy_memory.log.2026-10-19.gz"]
+        fresh = ["hy_memory.log.2026-10-21", "hy_memory.log.2026-10-22.gz"]
+        for name in expired + fresh:
+            (log_dir / name).write_text("x")
+        (log_dir / "hy_memory.log").write_text("active")
+        (log_dir / "unrelated.log").write_text("keep")
+
+        removed = log_setup.purge_expired_logs(
+            str(log_dir), 3, today=_dt.date(2026, 10, 23)
+        )
+
+        assert sorted(Path(p).name for p in removed) == sorted(expired)
+        assert {p.name for p in log_dir.iterdir()} == {
+            "hy_memory.log",
+            "unrelated.log",
+            *fresh,
+        }
+
+    def test_rotation_delete_list_includes_gzipped_files(self, tmp_path):
+        from pathlib import Path
+
+        log_dir = tmp_path / "logs"
+        log_dir.mkdir()
+        base = log_dir / "hy_memory.log"
+        base.write_text("active")
+        for day in ("01", "02", "03", "04"):
+            (log_dir / f"hy_memory.log.2026-10-{day}").write_text("x")
+        (log_dir / "hy_memory.log.2026-09-30.gz").write_text("x")
+
+        handler = log_setup.GzipAwareTimedRotatingFileHandler(
+            filename=str(base), when="midnight", backupCount=2
+        )
+        try:
+            to_delete = {Path(p).name for p in handler.getFilesToDelete()}
+        finally:
+            handler.close()
+
+        assert to_delete == {
+            "hy_memory.log.2026-09-30.gz",
+            "hy_memory.log.2026-10-01",
+            "hy_memory.log.2026-10-02",
+        }