superlocalmemory 3.8.3 → 3.8.5

package/src/superlocalmemory/server/unified_daemon.py

@@ -72,6 +72,22 @@ from superlocalmemory.learning.source_quality import (
     repair_historical_source_quality,
 )
 
+# D-02 fix: import at module load so the reference survives interpreter teardown.
+# A lazy import inside the shutdown path races the namespace-cleanup phase on
+# CPython and can raise ImportError: "cannot import name 'trigram_index'".
+try:
+    from superlocalmemory.learning import trigram_index as _trigram_index_mod
+except ImportError:  # pragma: no cover — defensive for stripped installs
+    _trigram_index_mod = None  # type: ignore[assignment]
+
+# D-03 fix: move this import to module top so the already-compiled .pyc object
+# is used at shutdown time. A lazy import of a source .py during interpreter
+# shutdown on macOS triggers TCC xattr checks and can raise PermissionError.
+try:
+    from superlocalmemory.hooks._outcome_common import _perf_log_flush as _perf_log_flush_fn
+except ImportError:  # pragma: no cover — defensive for stripped installs
+    _perf_log_flush_fn = None  # type: ignore[assignment]
+
 logger = logging.getLogger("superlocalmemory.unified_daemon")
 
 _DEFAULT_PORT = 8765
@@ -801,6 +817,29 @@ def _warm_spreading_activation(engine, runtime) -> bool:
                 sa.search(query_embedding, profile_id=active_pid, top_k=7)
         else:
             sa.search(query_embedding, profile_id=active_pid, top_k=7)
+        # v3.8.5: pre-load the per-profile graph-metrics cache (PageRank +
+        # community for every fact) here in the background, so the FIRST real
+        # recall does not pay the ~12k-row load on its hot path.  Observed as a
+        # large chunk of the 8-13s "cold first query" spike: the metrics cache
+        # was loading lazily during the user's first recall instead of at boot.
+        try:
+            sa._load_graph_metrics_cache(active_pid)
+        except Exception:
+            pass
+        # v3.8.5: warm the EntityGraphChannel's in-memory adjacency cache (ALL
+        # ~208K edges + entity maps + graph metrics, ~18 MB) at boot.  This was
+        # the dominant cause of the "cold first query" spike (8-13s): the cache
+        # loaded lazily on the FIRST real recall that routed to the entity
+        # channel, not during warmup.  Force it here via the channel's own
+        # search entry so it is warm before any user query.
+        try:
+            entity_channel = getattr(retr, "_entity", None)
+            if entity_channel is not None:
+                entity_channel.search(
+                    "memory graph adjacency warmup", active_pid, top_k=5,
+                )
+        except Exception:
+            pass
         logger.info(
             "Spreading-activation graph pre-warmed for profile %s", active_pid,
         )
@@ -1362,11 +1401,15 @@ async def lifespan(application: FastAPI):
         # full sort.  Without them full recall takes 7-10s on
         # >1M edges (the SpreadingActivation 4-UNION query disk-sorts every
         # node's neighbor list on each call).  With them: sub-second.
+        #
+        # Concurrency fix (v3.8.4): wrapped in memory_write() so the process
+        # write lock is acquired before the connection is opened.  Previously
+        # the raw sqlite3.connect() bypassed get_write_lock() and had no
+        # busy_timeout, which risked SQLITE_BUSY at daemon startup when other
+        # writers (hook / CLI) were already active.
         try:
-            import sqlite3 as _sqlite3
-            _idx_conn = _sqlite3.connect(str(_memory_db))
-            try:
-                _idx_conn.execute("PRAGMA journal_mode=WAL")
+            from superlocalmemory.storage.memory_write import memory_write as _mw
+            with _mw(_memory_db) as _idx_conn:
                 _idx_conn.execute(
                     "CREATE INDEX IF NOT EXISTS idx_edges_source_weight "
                     "ON graph_edges(profile_id, source_id, weight DESC)"
@@ -1383,10 +1426,6 @@ async def lifespan(application: FastAPI):
                     "CREATE INDEX IF NOT EXISTS idx_assoc_target_weight "
                     "ON association_edges(profile_id, target_fact_id, weight DESC)"
                 )
-            finally:
-                # CP-09: close even if an execute() raises, so the connection
-                # (and its file handle / shared DB lock) never leaks.
-                _idx_conn.close()
         except Exception as _idx_exc:
             logger.debug("SpreadingActivation covering indexes skipped: %s", _idx_exc)
 
@@ -1413,14 +1452,36 @@ async def lifespan(application: FastAPI):
         _embedding_warm = False
         def _warmup_embedder():
             global _embedding_warm
-            try:
-                embedder = getattr(retrieval_eng, '_embedder', None) if retrieval_eng else None
-                if embedder and hasattr(embedder, 'embed'):
-                    embedder.embed("warmup")
-                    _embedding_warm = True
-                    logger.info("Embedding worker pre-warmed (model resident, keep_alive=-1)")
-            except Exception as exc:
-                logger.warning("Embedding warmup failed: %s", exc)
+            import time as _t
+            # RETRY: the retrieval engine / embedder can be created lazily a
+            # moment AFTER this thread starts.  The old one-shot attempt often
+            # captured a None embedder and left _embedding_warm stuck False
+            # forever — so /health reported not-ready even though on-demand
+            # embeds worked.  A stuck not-ready can make an auto-start hook
+            # believe the daemon is down and spawn a DUPLICATE daemon (two
+            # processes writing memory.db → cross-process SQLITE_BUSY).  Poll
+            # until the real embedder (retrieval_eng._embedder, the same one
+            # recall uses) is available, warm it, and flip the flag.
+            for _attempt in range(240):  # ~120s max at 0.5s steps
+                try:
+                    _re = retrieval_eng or getattr(engine, '_retrieval_engine', None)
+                    embedder = getattr(_re, '_embedder', None) if _re else None
+                    if embedder is not None and hasattr(embedder, 'embed'):
+                        embedder.embed("warmup")
+                        _embedding_warm = True
+                        logger.info(
+                            "Embedding worker pre-warmed (model resident, "
+                            "keep_alive=-1)"
+                        )
+                        return
+                except Exception as exc:
+                    logger.debug("Embedding warmup attempt %d failed: %s",
+                                 _attempt, exc)
+                _t.sleep(0.5)
+            logger.warning(
+                "Embedding warmup did not complete after retries; /health "
+                "may report not-ready even though on-demand embeds work"
+            )
 
         def _warmup_recall():
             """v3.4.62: Fire a full recall after embedding warms up.
@@ -1462,6 +1523,20 @@ async def lifespan(application: FastAPI):
                 # v3.8: the --fast recalls above skip spreading activation; warm
                 # that channel directly so the first FULL recall is not cold.
                 _warm_spreading_activation(engine, profile_runtime)
+                # v3.8.5: the fast recalls above do NOT exercise the full ranking
+                # path or the agentic round, so the FIRST real user query that
+                # takes the full path paid an 8-13s cold cost (ranking model +
+                # graph-metrics load).  Fire ONE full (fast=False) recall here in
+                # the background so those load at boot, not on a user's query.
+                # Best-effort; never blocks readiness.
+                try:
+                    with profile_runtime.operation_nowait() as _fsnap:
+                        if _fsnap is not None:
+                            engine.recall(
+                                "memory recall performance", limit=5, fast=False,
+                            )
+                except Exception as _fexc:
+                    logger.debug("Full-path warmup skipped (non-fatal): %s", _fexc)
                 elapsed = round((_t.monotonic() - t0) * 1000)
                 logger.info(
                     "Recall engine pre-warmed in %dms", elapsed,
@@ -1479,6 +1554,11 @@ async def lifespan(application: FastAPI):
             """
             import time as _t
             from pathlib import Path as _P
+            if os.environ.get("SLM_DISABLE_VS_BACKFILL") == "1":
+                logger.info(
+                    "VS backfill disabled via SLM_DISABLE_VS_BACKFILL=1"
+                )
+                return
             for _ in range(120):
                 if _embedding_warm:
                     break
@@ -1510,7 +1590,42 @@ async def lifespan(application: FastAPI):
                         continue
                     if vs.count(pid) >= int(len(with_emb) * 0.98):
                         continue  # already complete — no-op
-                    n = vs.rebuild_from_facts(with_emb)
+                    # Fix: route each upsert through db._lock with cooperative
+                    # yield between facts.  Previously called
+                    # vs.rebuild_from_facts() which opens its own sqlite3
+                    # connection (bypassing db._lock), causing concurrent
+                    # backfill_missing_embeddings writes to hit SQLITE_BUSY
+                    # while holding db._lock — starving user writes for ~50 s.
+                    import os as _selfheal_os
+                    # Stage 2 (write-queue plan): batch upserts so the write lock
+                    # is acquired a FEW times with a clear release window (pause)
+                    # between batches — instead of thousands of rapid
+                    # re-acquisitions.  threading.RLock is NOT fair, so on a
+                    # fresh boot a waiting user /remember could be starved for
+                    # ~50 s behind the per-fact churn.  A real pause between
+                    # bounded batches guarantees the waiting user write wins the
+                    # lock promptly.  One-time backfill; steady state no-ops.
+                    _batch = max(1, int(
+                        _selfheal_os.environ.get("SLM_SELFHEAL_BATCH", "50")))
+                    _pause = max(0.0, float(
+                        _selfheal_os.environ.get("SLM_SELFHEAL_BATCH_PAUSE_S", "0.05")))
+                    n = 0
+                    for _i in range(0, len(with_emb), _batch):
+                        _chunk = with_emb[_i:_i + _batch]
+                        with db._lock:
+                            for _fact_id, _profile_id, _embedding in _chunk:
+                                try:
+                                    if vs.upsert(_fact_id, _profile_id, _embedding):
+                                        n += 1
+                                except Exception as _upsert_exc:
+                                    logger.warning(
+                                        "VS backfill[%s]: upsert failed for %s: %s",
+                                        pid, str(_fact_id)[:16], _upsert_exc,
+                                    )
+                        # Release window: let any waiting user write through
+                        # before grabbing the lock again.
+                        if _pause > 0:
+                            _t.sleep(_pause)
                     logger.info(
                         "VS backfill[%s]: indexed %d of %d embedded facts",
                         pid, n, len(with_emb),
@@ -2156,18 +2271,23 @@ async def lifespan(application: FastAPI):
         logger.warning("evolution cost-conn cache close failed: %s", exc)
 
     # Drop the trigram cache conn symmetrically.
+    # D-02: use module-level reference (_trigram_index_mod) to avoid re-importing
+    # during interpreter teardown when the learning namespace may be gone.
     try:
-        from superlocalmemory.learning import trigram_index as _ti
-        _ti._reset_cache_conn()
+        if _trigram_index_mod is not None:
+            _trigram_index_mod._reset_cache_conn()
     except Exception as exc:  # pragma: no cover — defensive
         logger.warning("trigram cache conn close failed: %s", exc)
 
     # Flush the perf-log fd explicitly (the atexit hook still fires
     # but explicit close here is cheap insurance against uvicorn
     # killing the process before atexit runs).
+    # D-03: use module-level reference (_perf_log_flush_fn) — a lazy import here
+    # triggers macOS TCC xattr checks on the source .py at shutdown time, causing
+    # PermissionError: "Operation not permitted: .../hooks/_outcome_common.py".
     try:
-        from superlocalmemory.hooks._outcome_common import _perf_log_flush
-        _perf_log_flush()
+        if _perf_log_flush_fn is not None:
+            _perf_log_flush_fn()
     except Exception as exc:  # pragma: no cover — defensive
         logger.warning("perf_log flush failed: %s", exc)
 
@@ -2473,7 +2593,33 @@ def _register_dashboard_routes(application: FastAPI) -> None:
 
     Extracted from api.py's create_app() to avoid duplicate MemoryEngine.
     """
-    from superlocalmemory.server.api import UI_DIR
+    from superlocalmemory.server.api import UI_DIR as _source_ui_dir
+
+    # D-04: Copy UI assets to the data dir at daemon startup to avoid macOS
+    # xattr/TCC PermissionError on source-tree files in editable installs.
+    # The data dir has no quarantine attributes; the copy is idempotent (same
+    # content from the same package) so concurrent daemon starts are safe.
+    # Falls back to the source path with a WARNING — never crashes the daemon.
+    import shutil as _shutil
+    _data_ui_dir = state_path("ui")
+    try:
+        _data_ui_dir.mkdir(parents=True, exist_ok=True)
+        if _source_ui_dir.is_dir():
+            _shutil.copytree(
+                str(_source_ui_dir),
+                str(_data_ui_dir),
+                dirs_exist_ok=True,  # idempotent; safe under concurrent starts
+            )
+        UI_DIR = _data_ui_dir
+    except Exception as _ui_copy_exc:
+        # D-04 fallback: source path. Logged as WARNING (not DEBUG) so operators
+        # can diagnose PermissionError on editable installs. Never silently hidden.
+        logger.warning(
+            "D-04: UI copy to data dir failed; serving from source path %s: %s",
+            _source_ui_dir,
+            _ui_copy_exc,
+        )
+        UI_DIR = _source_ui_dir
 
     # Rate limiting (graceful)
     try:
@@ -2523,7 +2669,8 @@ def _register_dashboard_routes(application: FastAPI) -> None:
         async def rate_limit_middleware(request, call_next):
             client_ip = request.client.host if request.client else "unknown"
             is_write = request.method in ("POST", "PUT", "DELETE", "PATCH")
-            loopback = client_ip in ("127.0.0.1", "::1")
+            from superlocalmemory.server.loopback import is_loopback as _is_lb_rl
+            loopback = _is_lb_rl(client_ip)
             if not loopback and is_rate_limit_exempt(client_ip):
                 return await call_next(request)
             if loopback:
@@ -2715,10 +2862,17 @@ def _register_dashboard_routes(application: FastAPI) -> None:
             "fallback; non-loopback writes will be rejected.", _auth_exc,
         )
         try:
-            from superlocalmemory.hooks.prewarm_auth import is_loopback as _is_lb
+            from superlocalmemory.server.loopback import is_loopback as _is_lb
         except Exception:
-            def _is_lb(h: str) -> bool:
-                return h in ("127.0.0.1", "::1", "localhost")
+            import ipaddress as _ipa_fb
+
+            def _is_lb(h: str) -> bool:  # type: ignore[misc]
+                if not h or h.lower() == "localhost":
+                    return bool(h)
+                try:
+                    return _ipa_fb.ip_address(h).is_loopback
+                except ValueError:
+                    return False
 
         @application.middleware("http")
         async def _failclosed_auth(request, call_next):
@@ -2734,9 +2888,14 @@ def _register_dashboard_routes(application: FastAPI) -> None:
                 )
             return await call_next(request)
 
-    # Static files
+    # Static files — UI_DIR is already the effective path (data-dir or source
+    # fallback) set by the D-04 copy block above; mkdir is a no-op for
+    # the data-dir path (already created) and guarded in the source-fallback case.
     from fastapi.staticfiles import StaticFiles
-    UI_DIR.mkdir(exist_ok=True)
+    try:
+        UI_DIR.mkdir(parents=True, exist_ok=True)
+    except Exception:
+        pass  # source path may not be writable; StaticFiles reads, not writes
     application.mount("/static", StaticFiles(directory=str(UI_DIR)), name="static")
 
     # Route modules
@@ -2975,9 +3134,14 @@ def _register_daemon_routes(application: FastAPI) -> None:
         }
         # request is None only for direct internal/test calls (no HTTP client),
         # which are trusted; over HTTP FastAPI always injects the real Request.
+        from superlocalmemory.server.loopback import is_loopback as _is_loopback_host
+        from superlocalmemory.server.write_identity import _TEST_ISOLATION_ALLOWED
+
         client_host = request.client.host if (request and request.client) else ""
-        _trusted = request is None or client_host in (
-            "127.0.0.1", "::1", "localhost", "testclient",
+        _trusted = (
+            request is None
+            or _is_loopback_host(client_host)
+            or (client_host == "testclient" and _TEST_ISOLATION_ALLOWED)
         )
         if not _trusted:
             return public
@@ -4007,7 +4171,12 @@ def start_server(port: int = _DEFAULT_PORT) -> None:
             "reachable from the network but SLM_REQUIRE_CREDENTIALS is not set "
             "and API-key auth may be off. A remote caller could write without "
             "credentials. Set SLM_REQUIRE_CREDENTIALS=1 and configure an API "
-            "key before exposing this instance.", bind_host,
+            "key before exposing this instance. "
+            "NOTE (issue #90): SLM 3.8.4 fixes IPv4-mapped loopback address "
+            "normalization — if curl/dashboard writes fail with 403 from a "
+            "container, upgrade to 3.8.4. Immediate workaround: set "
+            "SLM_DAEMON_HOST=127.0.0.1 (IPv4 only) or configure an API key.",
+            bind_host,
         )
     listener = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
     # This handles a just-closed connection in TIME_WAIT.  It is safe only

package/src/superlocalmemory/server/write_identity.py

@@ -18,6 +18,8 @@ from typing import Any
 
 from fastapi import HTTPException, Request
 
+from superlocalmemory.server.loopback import is_loopback as _is_loopback_host
+
 
 # H-04 (3.7.9): the test-client auth bypass must be impossible in a real daemon
 # even if SLM_TEST_ISOLATION=1 leaks into a production environment. Anchor it to
@@ -89,7 +91,7 @@ def require_write_actor(
         # missing peer address (e.g. behind a proxy that strips it) must not be
         # trusted just because an install token is presented. require_http_mutation_actor
         # already excludes "" from its loopback set; keep the two paths consistent.
-        if is_test_client or client_host in ("127.0.0.1", "::1", "localhost"):
+        if is_test_client or _is_loopback_host(client_host):
             from superlocalmemory.core.engine_ingestion import local_trusted_actor_id
 
             return local_trusted_actor_id(actor_kind)
@@ -103,7 +105,15 @@ def require_write_actor(
         ).hexdigest()
         return f"api-key:{actor_kind}:{fingerprint}"
 
-    raise HTTPException(403, detail="Authenticated write capability required")
+    raise HTTPException(
+        403,
+        detail=(
+            "Write rejected: this origin requires an API key (X-SLM-API-Key). "
+            "The install token (X-Install-Token) is accepted only from loopback "
+            "(127.x.x.x / ::1 / localhost). If calling from a container or "
+            "remote host, configure SLM_API_KEY and present it as X-SLM-API-Key."
+        ),
+    )
 
 
 def require_http_mutation_actor(
@@ -134,7 +144,7 @@ def require_http_mutation_actor(
 
     client_host = request.client.host if request.client else ""
     is_test_client = client_host == "testclient" and _TEST_ISOLATION_ALLOWED
-    loopback = client_host in ("127.0.0.1", "::1", "localhost")
+    loopback = _is_loopback_host(client_host)
     # H-01: the loopback trusted-actor bypass is suppressed when the operator
     # opts into SLM_REQUIRE_CREDENTIALS; in-process tests keep the bypass.
     if is_test_client or (loopback and not _REQUIRE_CREDENTIALS):
@@ -153,7 +163,15 @@ def require_http_mutation_actor(
             ).hexdigest()
             return f"mesh-secret:{fingerprint}"
 
-    raise HTTPException(403, detail="Authenticated mutation capability required")
+    raise HTTPException(
+        403,
+        detail=(
+            "Mutation rejected: present one of X-SLM-Daemon-Capability, "
+            "X-Install-Token (loopback only), or X-SLM-API-Key. "
+            "Uncredentialed writes are accepted only from loopback. "
+            "See SLM_API_KEY in the deployment guide for network access."
+        ),
+    )
 
 
 def authenticated_request_actor(

package/src/superlocalmemory/storage/database.py

@@ -18,6 +18,7 @@ from pathlib import Path
 from types import ModuleType
 from typing import Any, Generator
 
+from superlocalmemory.storage.write_lock import get_write_lock
 from superlocalmemory.storage.models import (
     AtomicFact, CanonicalEntity, ConsolidationAction, ConsolidationActionType,
     EdgeType, EntityAlias, EntityProfile, FactType, GraphEdge,
@@ -140,7 +141,17 @@ class DatabaseManager:
     def __init__(self, db_path: str | Path) -> None:
         self.db_path = Path(db_path)
         self.db_path.parent.mkdir(parents=True, exist_ok=True)
-        self._lock = threading.Lock()
+        # Shared write-serialisation lock for this db_path.
+        # get_write_lock() returns the SAME RLock for every caller that
+        # passes the same resolved path, so DatabaseManager, VectorStore,
+        # adapter_base, consolidation, and all other in-process writers
+        # all share ONE lock → zero cross-connection SQLite WAL contention.
+        #
+        # RLock (re-entrant) is required: the self-heal backfill pattern
+        #   with db._lock:        # acquires write lock (count: 1→2)
+        #       vs.upsert(...)    # re-acquires same lock (count: 2→3)
+        # is safe because the same thread re-enters the RLock.
+        self._lock = get_write_lock(self.db_path)
         # Transaction connections are thread-affine in sqlite3. A manager is
         # shared across HTTP, materializer, and worker threads, so a process-
         # global connection slot lets another thread accidentally execute on
@@ -155,6 +166,20 @@ class DatabaseManager:
             conn.execute(f"PRAGMA busy_timeout={_BUSY_TIMEOUT_MS}")  # FIRST — so WAL pragma below uses configured timeout
             conn.execute("PRAGMA journal_mode=WAL")
             conn.execute("PRAGMA foreign_keys=ON")
+            # Fix D: synchronous=NORMAL is safe under WAL — the WAL ensures
+            # atomicity independently of the fsync level.  Removing the
+            # fsync-on-every-commit penalty halves write latency under the
+            # 14 000+ edge-write bursts the materialiser produces per pass.
+            # Trade-off: on a power loss between COMMIT and WAL-checkpoint
+            # the last committed write to memory.db may be lost.  Acceptable
+            # for an LLM memory system; NOT acceptable for financial records.
+            conn.execute("PRAGMA synchronous=NORMAL")
+            # Fix D: reduce WAL auto-checkpoint from the default 1000 to 400
+            # frames.  Smaller checkpoints run more frequently and complete
+            # faster, preventing the WAL file growing unboundedly during
+            # high-ingestion bursts (which triggered checkpoint-starvation
+            # amplifying the lock storm).
+            conn.execute("PRAGMA wal_autocheckpoint=400")
             conn.commit()
         finally:
             conn.close()
@@ -231,15 +256,21 @@ class DatabaseManager:
                 self._txn_state.conn = None
                 conn.close()
 
-    def execute(self, sql: str, params: tuple[Any, ...] = ()) -> list[sqlite3.Row]:
-        """Execute SQL with automatic retry on SQLITE_BUSY.
+    # DML prefixes that require the single-writer lock when executed outside
+    # a transaction() context.  Checked case-insensitively against the first
+    # word of the stripped SQL statement.
+    _DML_PREFIXES: frozenset[str] = frozenset({
+        "INSERT", "UPDATE", "DELETE", "REPLACE", "UPSERT",
+        "CREATE", "DROP", "ALTER",
+    })
 
-        Uses shared conn inside transaction, else per-call with retry.
-        """
-        transaction_conn = getattr(self._txn_state, "conn", None)
-        if transaction_conn is not None:
-            return transaction_conn.execute(sql, params).fetchall()
+    def _execute_one(self, sql: str, params: tuple[Any, ...]) -> list[sqlite3.Row]:
+        """Open a per-call connection, execute, commit, close — with retry.
 
+        Never called when inside a transaction() context (that path uses the
+        context's existing connection directly).  Factored out of execute() so
+        the RLock acquisition logic stays in one place.
+        """
         last_error: Exception | None = None
         for attempt in range(_MAX_RETRIES):
             conn = self._connect()
@@ -264,6 +295,42 @@ class DatabaseManager:
         logger.warning("DB operation failed after %d retries: %s", _MAX_RETRIES, last_error)
         raise last_error  # type: ignore[misc]
 
+    def execute(self, sql: str, params: tuple[Any, ...] = ()) -> list[sqlite3.Row]:
+        """Execute SQL with automatic retry on SQLITE_BUSY.
+
+        Fix B — single-writer serialisation:
+        • Inside a transaction(): uses the existing connection directly (no
+          lock acquisition — the transaction() context manager already holds
+          _lock for the duration of the whole transaction).
+        • Outside a transaction(), for DML (INSERT/UPDATE/DELETE/…): acquires
+          _lock before opening a per-call connection.  This ensures that
+          concurrent callers — including background workers that bypass
+          transaction() entirely — do not race at the SQLite WAL layer.
+        • Outside a transaction(), for SELECTs: no lock needed — WAL mode
+          allows concurrent readers without stalling writers.
+
+        ORDERING INVARIANT: the _txn_state.conn check MUST come before the
+        lock acquisition.  Reversing the order would deadlock threads that
+        call execute() from inside transaction() because threading.RLock is
+        re-entrant per-thread but a re-entering thread inside transaction()
+        would still try to re-acquire here (lock is already held by the same
+        thread, so RLock re-enters safely — but the old threading.Lock would
+        have deadlocked; that is exactly why we changed to RLock).
+        """
+        # Fast path: already inside a transaction — use its connection directly.
+        transaction_conn = getattr(self._txn_state, "conn", None)
+        if transaction_conn is not None:
+            return transaction_conn.execute(sql, params).fetchall()
+
+        # Determine if this is a write operation that needs serialisation.
+        first_word = sql.strip().upper().split(None, 1)[0] if sql.strip() else ""
+        if first_word in self._DML_PREFIXES:
+            with self._lock:
+                return self._execute_one(sql, params)
+        else:
+            # Read-only path: concurrent reads are safe in WAL mode.
+            return self._execute_one(sql, params)
+
     def store_memory(self, record: MemoryRecord) -> str:
         """Persist a raw memory record. Returns memory_id."""
         _scope = getattr(record, 'scope', None) or 'personal'
@@ -1306,17 +1373,40 @@ class DatabaseManager:
         )
         return [dict(r) for r in rows]
 
-    def cleanup_activation_cache(self) -> int:
-        """Delete expired cache entries. Returns count deleted."""
-        before = self.execute(
-            "SELECT COUNT(*) AS c FROM activation_cache "
-            "WHERE expires_at < datetime('now')"
-        )
-        count = int(before[0]["c"]) if before else 0
-        self.execute(
-            "DELETE FROM activation_cache WHERE expires_at < datetime('now')"
-        )
-        return count
+    def cleanup_activation_cache(
+        self, batch_size: int = 5000, max_batches: int = 500,
+    ) -> int:
+        """Delete expired activation_cache rows in bounded batches.
+
+        Wired into MaintenanceScheduler.  Historically NEITHER cleanup path was
+        ever called, so activation_cache grew without bound — observed 83,518
+        rows on a real DB, all expired, oldest ~3.5 months old.  That bloats the
+        table and its ``idx_actcache_expires`` index and slows every cache
+        INSERT OR REPLACE / lookup.
+
+        Batched so clearing a large backlog never holds the write lock for one
+        long DELETE: each batch commits and yields, letting remember/materialize
+        writers interleave.  ``idx_actcache_expires`` makes the predicate
+        index-backed.  Steady state (30-min cycle) deletes only one cycle's
+        worth, so the loop exits after a single small batch.
+        """
+        total_deleted = 0
+        for _ in range(max_batches):
+            remaining = self.execute(
+                "SELECT COUNT(*) AS c FROM activation_cache "
+                "WHERE expires_at < datetime('now')"
+            )
+            n = int(remaining[0]["c"]) if remaining else 0
+            if n <= 0:
+                break
+            self.execute(
+                "DELETE FROM activation_cache WHERE cache_id IN ("
+                "  SELECT cache_id FROM activation_cache "
+                "  WHERE expires_at < datetime('now') LIMIT ?)",
+                (batch_size,),
+            )
+            total_deleted += min(n, batch_size)
+        return total_deleted
 
     def store_fact_importance(self, entry: dict) -> None:
         """Persist fact importance scores."""