superlocalmemory 3.4.31 → 3.4.33

package/CHANGELOG.md

@@ -10,6 +10,62 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ---
 
+## [3.4.33] - 2026-04-25
+
+Fix: daemon leaked SQLite connections to learning.db via bandit threadlocals.
+
+### Fixed
+- **Bandit threadlocal connection leak.** `reward_proxy.settle_stale_plays`
+  creates a `ContextualBandit` that opens a threadlocal connection via
+  `_conn_for`. When called from `asyncio.to_thread` (bandit_loops.py,
+  every 60 s), each thread-pool thread kept its connection open for the
+  process lifetime. Over 24 h this accumulated 12+ leaked file descriptors
+  and ~100 MB of wasted SQLite page-cache RAM. New
+  `bandit.close_threadlocal_conn()` function, called in the
+  `settle_stale_plays` finally block, ensures pool threads release their
+  connections immediately.
+- **Corrected embedding worker memory comment.** The `~200MB footprint`
+  note was written for `all-MiniLM-L6-v2`; the default model
+  `nomic-ai/nomic-embed-text-v1.5` uses ~1.1 GB via ONNX.
+
+---
+
+## [3.4.32] - 2026-04-24
+
+Fix: concurrent remembers no longer block recalls on the shared embedder.
+
+### Fixed
+- **Daemon `/remember` is now async by default.** Writes to the pending
+  queue in under 100 ms and returns a `pending_id`; a background thread
+  drains the queue in the background. Previously, the synchronous
+  `engine.store()` on the FastAPI event loop could block `/search` and
+  `/health` for 30+ seconds while the single embedder worker processed a
+  large write. Under concurrent load the daemon could appear hung.
+- **Materializer yields to active recalls.** While any `/search` is in
+  flight the drainer sleeps between items, so user-initiated recalls
+  always get the embedder first.
+- **MCP remember tool simplified.** Writes to `pending.db` and returns;
+  the daemon's materializer completes the pipeline. Removes the
+  redundant in-process `pool.store` background task that previously
+  contended with `/search`.
+- **`pool_store` returns `["pending:<id>"]`** when the daemon is async,
+  keeping a stable identifier for callers without blocking on the
+  embedder.
+
+### Added
+- `?wait=true` query parameter on `POST /remember` for callers that
+  need synchronous behaviour and real `fact_ids` in the response.
+- `superlocalmemory.core.recall_gate` module — shared counter that lets
+  the materializer detect in-flight recalls and yield priority.
+
+### Migration notes
+- **No action required.** Existing clients continue to work; the
+  response shape is compatible (`ok`, `count` still present). Scripts
+  that depended on `fact_ids` to validate the write should switch to
+  `pending_id` or pass `?wait=true` to opt in to the legacy behaviour.
+
+---
+
 ## [3.4.31] - 2026-04-24
 
 Dashboard truth, memory vs fact clarity, and self-cleaning pending queue.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "superlocalmemory",
-  "version": "3.4.31",
+  "version": "3.4.33",
   "description": "Information-geometric agent memory with mathematical guarantees. 4-channel retrieval, Fisher-Rao similarity, zero-LLM mode, EU AI Act compliant. Works with Claude, Cursor, Windsurf, and 17+ AI tools.",
   "keywords": [
     "ai-memory",

package/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "superlocalmemory"
-version = "3.4.31"
+version = "3.4.33"
 description = "Information-geometric agent memory with mathematical guarantees"
 readme = "README.md"
 license = {text = "AGPL-3.0-or-later"}

package/src/superlocalmemory/__init__.py

@@ -1,3 +1,3 @@
 """SuperLocalMemory — information-geometric agent memory."""
 
-__version__ = "3.4.31"
+__version__ = "3.4.33"

package/src/superlocalmemory/core/embedding_worker.py

@@ -63,7 +63,7 @@ def _load_embedding_model(name: str) -> tuple:
     """
     from sentence_transformers import SentenceTransformer
 
-    # Tier 1: ONNX (stable memory, ~200MB footprint)
+    # Tier 1: ONNX (stable memory; ~1.1 GB for nomic-embed-text-v1.5)
     try:
         m = SentenceTransformer(name, backend="onnx", trust_remote_code=True)
         return m, "onnx"

package/src/superlocalmemory/core/recall_gate.py

@@ -0,0 +1,36 @@
+# Copyright (c) 2026 Varun Pratap Bhardwaj / Qualixar
+# Licensed under AGPL-3.0-or-later - see LICENSE file
+# Part of SuperLocalMemory V3 | https://qualixar.com | https://varunpratap.com
+
+"""v3.4.32: Recall-in-flight counter used to give /search priority over the
+pending materializer.
+
+Every recall handler calls ``begin_recall()`` on entry and ``end_recall()``
+in a finally block. The pending-memory materializer thread polls
+``in_flight()`` and sleeps while any recall is active, so the shared
+embedder worker never serves a materialization ahead of a user-initiated
+recall.
+"""
+from __future__ import annotations
+
+import threading
+
+_lock = threading.Lock()
+_active = 0
+
+
+def begin_recall() -> None:
+    global _active
+    with _lock:
+        _active += 1
+
+
+def end_recall() -> None:
+    global _active
+    with _lock:
+        _active = max(0, _active - 1)
+
+
+def in_flight() -> int:
+    with _lock:
+        return _active

package/src/superlocalmemory/learning/bandit.py

@@ -176,6 +176,23 @@ def _conn_for(db_path: Path) -> sqlite3.Connection:
     return conn
 
 
+def close_threadlocal_conn() -> None:
+    """Close the threadlocal bandit connection on the calling thread.
+
+    v3.4.33: background callers (asyncio.to_thread pool threads) MUST call
+    this after finishing bandit work.  Without it, each pool thread keeps a
+    leaked connection to learning.db for the process lifetime — observed as
+    12+ open file descriptors and ~100 MB wasted page-cache RAM.
+    """
+    if _holder.conn is not None:
+        try:
+            _holder.conn.close()
+        except sqlite3.Error:  # pragma: no cover
+            pass
+        _holder.conn = None
+        _holder.path = None
+
+
 def _now_iso() -> str:
     return datetime.now(timezone.utc).isoformat(timespec="seconds")
 
@@ -520,6 +537,7 @@ def retention_sweep(
 __all__ = (
     "BanditChoice",
     "ContextualBandit",
+    "close_threadlocal_conn",
     "compute_stratum",
     "current_time_bucket",
     "retention_sweep",

package/src/superlocalmemory/learning/reward_proxy.py

@@ -317,8 +317,13 @@ def settle_stale_plays(
                 memory_conn.close()
             except sqlite3.Error:  # pragma: no cover
                 pass
-        # Don't close a caller-owned bandit instance.
-        _ = owns_bandit
+        # v3.4.33: close the threadlocal bandit connection so pool threads
+        # from asyncio.to_thread don't leak file descriptors to learning.db.
+        try:
+            from superlocalmemory.learning.bandit import close_threadlocal_conn
+            close_threadlocal_conn()
+        except Exception:  # pragma: no cover — defensive
+            pass
 
     return settled
 

package/src/superlocalmemory/mcp/_pool_adapter.py

@@ -110,12 +110,24 @@ def pool_recall(query: str, limit: int = 10, **_: Any) -> PoolRecallResponse:
 
 
 def pool_store(content: str, metadata: dict | None = None) -> list[str]:
-    """Call pool.store and return the fact id list.
+    """Call pool.store and return fact id list (or pending tracker).
 
-    Raises :class:`PoolError` on worker death or any non-ok envelope.
+    v3.4.32: the daemon /remember endpoint is async by default — it
+    returns ``pending_id`` and queues the write. We surface this to
+    callers as ``["pending:<id>"]`` so they have a stable identifier
+    without blocking the remember on the embedder worker.
+
+    Legacy synchronous path (``?wait=true``) still returns real
+    ``fact_ids``. Worker death raises :class:`PoolError`.
     """
     raw = _pool().store(content=content, metadata=metadata or {})
     _unwrap_error(raw, "store")
-    if isinstance(raw, dict):
-        return list(raw.get("fact_ids", []))
+    if not isinstance(raw, dict):
+        return []
+    fact_ids = raw.get("fact_ids")
+    if fact_ids:
+        return list(fact_ids)
+    pending_id = raw.get("pending_id")
+    if pending_id is not None:
+        return [f"pending:{pending_id}"]
     return []

package/src/superlocalmemory/mcp/tools_core.py

@@ -113,12 +113,11 @@ def register_core_tools(server, get_engine: Callable) -> None:
         """
         import asyncio
         try:
-            # V3.3.27: Store-first pattern — write to pending.db immediately
-            # (<100ms), then process through full pipeline in background.
-            # This eliminates the 30-40s blocking that Mode B users experience.
-            # Pending memories are auto-processed on next engine.initialize()
-            # or by the daemon's background loop.
-            from superlocalmemory.cli.pending_store import store_pending, mark_done
+            # v3.4.32: Store-first pattern. Write to pending.db and return
+            # immediately. The daemon's pending-materializer thread drains
+            # the queue with recall priority, so concurrent MCP remembers
+            # no longer contend with /search on the shared embedder.
+            from superlocalmemory.cli.pending_store import store_pending
 
             pending_id = store_pending(content, tags=tags, metadata={
                 "project": project,
@@ -127,39 +126,13 @@ def register_core_tools(server, get_engine: Callable) -> None:
                 "session_id": session_id,
             })
 
-            # Fire-and-forget: process in background thread
-            async def _process_in_background():
-                try:
-                    from superlocalmemory.core.worker_pool import WorkerPool
-                    pool = WorkerPool.shared()
-                    result = await asyncio.to_thread(
-                        pool.store, content, metadata={
-                            "tags": tags, "project": project,
-                            "importance": importance, "agent_id": agent_id,
-                            "session_id": session_id,
-                        },
-                    )
-                    if result.get("ok"):
-                        mark_done(pending_id)
-                        _emit_event("memory.created", {
-                            "content_preview": content[:80],
-                            "agent_id": agent_id,
-                            "fact_count": result.get("count", 0),
-                        }, source_agent=agent_id)
-                except Exception as _bg_exc:
-                    logger.warning(
-                        "Background store failed (pending_id=%s): %s",
-                        pending_id, _bg_exc,
-                    )
-
-            asyncio.create_task(_process_in_background())
-
             return {
                 "success": True,
                 "fact_ids": [f"pending:{pending_id}"],
                 "count": 1,
                 "pending": True,
-                "message": "Stored to pending — processing in background.",
+                "pending_id": pending_id,
+                "message": "Stored — facts will appear in the dashboard shortly.",
             }
         except Exception as exc:
             logger.exception("remember failed")

package/src/superlocalmemory/server/routes/memories.py

@@ -398,7 +398,12 @@ async def get_graph(
 
 @router.post("/api/search")
 async def search_memories(request: Request, body: SearchRequest):
-    """Semantic search via subprocess worker pool (memory-isolated)."""
+    """Semantic search via subprocess worker pool (memory-isolated).
+
+    v3.4.32: marks recall in-flight so the pending materializer yields.
+    """
+    from superlocalmemory.core.recall_gate import begin_recall, end_recall
+    begin_recall()
     try:
         from superlocalmemory.core.worker_pool import WorkerPool
         pool = WorkerPool.shared()
@@ -435,6 +440,8 @@ async def search_memories(request: Request, body: SearchRequest):
 
     except Exception as e:
         raise HTTPException(status_code=500, detail=f"Search error: {str(e)}")
+    finally:
+        end_recall()
 
 
 @router.get("/api/clusters")

package/src/superlocalmemory/server/unified_daemon.py

@@ -66,6 +66,20 @@ class ObserveRequest(BaseModel):
     content: str
 
 
+# ---------------------------------------------------------------------------
+# v3.4.32: Recall-priority gate for the pending materializer.
+# All /remember writes go to pending.db and return fast; a background
+# thread drains pending while yielding to any in-flight /search.
+# See ``superlocalmemory.core.recall_gate``.
+# ---------------------------------------------------------------------------
+
+from superlocalmemory.core.recall_gate import (
+    begin_recall as _begin_recall,
+    end_recall as _end_recall,
+    in_flight as _recalls_in_flight,
+)
+
+
 # ---------------------------------------------------------------------------
 # Observation debounce buffer (migrated from daemon.py)
 # ---------------------------------------------------------------------------
@@ -949,6 +963,8 @@ def _register_daemon_routes(application: FastAPI) -> None:
         if not effective_sid:
             import time as _t
             effective_sid = f"http:{int(_t.time() * 1000)}"
+        # v3.4.32: mark recall in-flight so the pending materializer pauses
+        _begin_recall()
         try:
             response = engine.recall(
                 search_query, limit=limit, session_id=effective_sid,
@@ -1006,18 +1022,47 @@ def _register_daemon_routes(application: FastAPI) -> None:
             }
         except Exception as exc:
             raise HTTPException(500, detail=str(exc))
+        finally:
+            _end_recall()
 
     @application.post("/remember")
-    async def remember(req: RememberRequest):
+    async def remember(req: RememberRequest, wait: bool = False):
+        """v3.4.32: Async by default — writes to pending.db, returns pending_id
+        in <100ms. Materializer thread drains at low priority, yielding to
+        /search. Pass ``?wait=true`` for legacy synchronous behavior (blocks
+        on the embedder until facts are written).
+        """
         _update_activity()
         engine = _get_engine_or_503()
+
+        if wait:
+            try:
+                metadata = {"tags": req.tags} if req.tags else {}
+                extra = getattr(req, "metadata", None)
+                if isinstance(extra, dict):
+                    metadata.update(extra)
+                fact_ids = engine.store(req.content, metadata=metadata)
+                return {"ok": True, "fact_ids": fact_ids, "count": len(fact_ids)}
+            except Exception as exc:
+                raise HTTPException(500, detail=str(exc))
+
         try:
-            metadata = {"tags": req.tags} if req.tags else {}
+            from superlocalmemory.cli.pending_store import store_pending
+            meta = {}
+            if req.tags:
+                meta["tags"] = req.tags
             extra = getattr(req, "metadata", None)
             if isinstance(extra, dict):
-                metadata.update(extra)
-            fact_ids = engine.store(req.content, metadata=metadata)
-            return {"ok": True, "fact_ids": fact_ids, "count": len(fact_ids)}
+                meta.update(extra)
+            pending_id = store_pending(
+                req.content, tags=req.tags or "", metadata=meta,
+            )
+            return {
+                "ok": True,
+                "pending_id": pending_id,
+                "status": "queued",
+                "note": "materialized async; pass ?wait=true for legacy sync",
+            }
         except Exception as exc:
             raise HTTPException(500, detail=str(exc))
 
@@ -1189,6 +1234,70 @@ def _start_memory_watchdog() -> None:
     logger.info("Memory watchdog started (limit: %d MB per worker)", MAX_WORKER_MB)
 
 
+_materializer_stop = threading.Event()
+_materializer_thread: threading.Thread | None = None
+
+
+def _start_pending_materializer() -> None:
+    """Background thread: drains pending.db, yields to active /search calls.
+
+    Poll loop:
+    1. Fetch up to 5 pending rows.
+    2. For each row: if any /search is in flight, sleep 500ms (yield priority).
+    3. Call engine.store(), mark_done or mark_failed.
+    4. Sleep 2s between polls when idle (empty queue).
+    """
+    global _materializer_thread
+
+    def _loop():
+        from superlocalmemory.cli.pending_store import (
+            get_pending, mark_done, mark_failed,
+        )
+        while not _materializer_stop.is_set():
+            try:
+                engine = _engine  # may be None briefly at startup
+                if engine is None:
+                    time.sleep(2.0)
+                    continue
+                pending = get_pending(limit=5)
+                if not pending:
+                    time.sleep(2.0)
+                    continue
+                for item in pending:
+                    if _materializer_stop.is_set():
+                        break
+                    # Yield to recalls: wait until none in flight
+                    waits = 0
+                    while _recalls_in_flight() > 0 and waits < 60:
+                        time.sleep(0.5)
+                        waits += 1
+                    try:
+                        import json as _json
+                        md_str = item.get("metadata") or "{}"
+                        try:
+                            md = _json.loads(md_str)
+                        except Exception:
+                            md = {}
+                        if item.get("tags"):
+                            md.setdefault("tags", item["tags"])
+                        engine.store(item["content"], metadata=md)
+                        mark_done(item["id"])
+                    except Exception as exc:
+                        logger.warning(
+                            "Pending %d failed: %s", item["id"], exc,
+                        )
+                        mark_failed(item["id"], str(exc))
+            except Exception as exc:
+                logger.warning("materializer loop error: %s", exc)
+                time.sleep(5.0)
+
+    _materializer_thread = threading.Thread(
+        target=_loop, daemon=True, name="pending-materializer",
+    )
+    _materializer_thread.start()
+    logger.info("Pending materializer started (recall-priority)")
+
+
 def start_server(port: int = _DEFAULT_PORT) -> None:
     """Start the unified daemon. Blocks until stopped."""
     global _start_time
@@ -1223,6 +1332,9 @@ def start_server(port: int = _DEFAULT_PORT) -> None:
     # v3.4.7: Start memory watchdog to prevent runaway workers
     _start_memory_watchdog()
 
+    # v3.4.32: Continuous pending-queue materializer with recall priority.
+    _start_pending_materializer()
+
     log_dir = Path.home() / ".superlocalmemory" / "logs"
     log_dir.mkdir(parents=True, exist_ok=True)