fix(memory): flush memory queue on graceful shutdown to prevent loss (#4181)

* fix(memory): bounded shutdown flush via MemoryManager.shutdown_flush

Re-applies the memory-queue shutdown drain on top of the pluggable
MemoryManager abstraction (#4122): the old top-level MemoryUpdateQueue
singleton is gone, so the drain is now a backend contract instead of
host code reaching into the queue.

- MemoryManager ABC: shutdown_flush(timeout) -> bool. Every backend
  implements a bounded graceful-shutdown drain.
- DeerMem: queue.flush_sync (daemon-thread + Event.wait hard timeout
  for the uninterruptible sync LLM call; joins an in-flight worker
  first so contexts a debounce Timer already pulled out are not lost on
  exit; skips inter-item sleep on the drain path; per-item
  succeeded/failed count), exposed via shutdown_flush.
- noop: shutdown_flush is a clean no-op success.
- Gateway lifespan: call get_memory_manager().shutdown_flush(timeout)
  after channels/scheduler stop, via asyncio.to_thread, try/except
  bounded. No host-level pending/processing guard -- the backend
  short-circuits on an idle buffer, so the host cannot "forget" the
  in-flight case (structurally eliminates the guard race flagged on the
  prior revision).
- shutdown_flush_timeout_seconds added to the shared MemoryConfig
  (host-owned lifecycle budget, default 30, 1-300) + exposed on
  MemoryConfigResponse and the embedded client; config_version 25 -> 26.

Tests: queue flush_sync (7), lifespan drain incl. False-branch caplog
assertion + disabled gate (3), ABC contract noop/deermem (3).

* fix(chart): gateway grace period so memory drain is not SIGKILLed

K8s defaults terminationGracePeriodSeconds to 30s, shorter than the
Gateway's graceful-shutdown work (channel stop ~5s + memory queue drain
default 30s). Without an explicit grace period, K8s SIGKILLs the memory
drain mid-flight and silently re-introduces the loss shutdown_flush is
fixing (flagged on the prior revision).

- gateway pod: terminationGracePeriodSeconds (default 45, configurable).
- gateway container: preStop sleep (default 5, 0 disables) so the
  Service/ingress deregisters the pod before SIGTERM begins the drain.
- values.yaml + README: both configurable; README documents that the
  grace period must track memory.shutdown_flush_timeout_seconds.

* docs(memory): document shutdown_flush_timeout_seconds + lifespan drain

Add the host-shared field to the memory config list and Config Schema
summary in backend/AGENTS.md, noting the lifespan drain and the K8s
grace-period relationship.

* fix(chart): bump embedded config_version to 26

The chart's embedded `config:` block (values.yaml + README example) still
had config_version: 25 after commit f3ca8e9f raised config.example.yaml to
26, failing the validate-chart config_version drift check. Bump both to 26.

backend/AGENTS.md

@@ -596,6 +596,7 @@ Focused regression coverage for the updater lives in `backend/tests/test_memory_
 - `mode` - Operation mode: `middleware` (default passive background extraction) or `tool` (experimental model-driven memory tools). Modes are mutually exclusive.
 - `storage_path` - Path to memory.json (absolute path opts out of per-user isolation)
 - `debounce_seconds` - Wait time before processing (default: 30)
+- `shutdown_flush_timeout_seconds` - Host-shared hard budget (seconds) to drain the memory backend's pending-update buffer on Gateway graceful shutdown (default: 30; 1–300). Each pending item does one LLM call, so large IM batches may need more. The Gateway lifespan calls `MemoryManager.shutdown_flush(timeout)` after channels/scheduler stop; the backend short-circuits on an idle buffer, so the host calls it unconditionally (no pending/processing gate). Must fit inside the pod's K8s `terminationGracePeriodSeconds` (gateway Helm chart sets this; default 45s) or K8s SIGKILLs the drain mid-flight.
 - `model_name` - LLM for updates (null = default model)
 - `max_facts` / `fact_confidence_threshold` - Fact storage limits (100 / 0.7)
 - `max_injection_tokens` - Token limit for prompt injection (2000)
@@ -732,7 +733,7 @@ Config is env-driven like the others — `MonocleTracingConfig`, built in `get_t
 - `title` - Auto-title generation (enabled, max_words, max_chars, model_name; null model_name uses fast local fallback, explicit model_name uses the prompt_template LLM path)
 - `summarization` - Context summarization (enabled, trigger conditions, keep policy)
 - `subagents.enabled` - Master switch for subagent delegation
-- `memory` - Memory system (enabled, storage_path, debounce_seconds, model_name, max_facts, fact_confidence_threshold, injection_enabled, max_injection_tokens, staleness_review_enabled, staleness_age_days, staleness_min_candidates, staleness_max_removals_per_cycle, staleness_protected_categories)
+- `memory` - Memory system (enabled, storage_path, debounce_seconds, shutdown_flush_timeout_seconds, model_name, max_facts, fact_confidence_threshold, injection_enabled, max_injection_tokens, staleness_review_enabled, staleness_age_days, staleness_min_candidates, staleness_max_removals_per_cycle, staleness_protected_categories)
 
 **`extensions_config.json`**:
 - `mcpServers` - Map of server name → config (enabled, type, command, args, env, url, headers, oauth, description, `routing`, `tools`, `tool_call_timeout`). `routing.mode="prefer"` emits `<mcp_routing_hints>` prompt guidance; if `tool_search` defers the hinted tool, `McpRoutingMiddleware` can also auto-promote matching deferred schemas before the model call. It does not hard-disable other tools.

backend/app/gateway/app.py

@@ -301,6 +301,43 @@ async def lifespan(app: FastAPI) -> AsyncGenerator[None, None]:
             except Exception:
                 logger.exception("Failed to stop scheduled task service")
 
+        # Drain the memory backend's pending-update buffer before the worker
+        # exits (best-effort, bounded). IM channels and the scheduler are
+        # already stopped above, so no new IM/scheduler updates arrive during
+        # the drain; the LangGraph runtime / in-flight HTTP requests can still
+        # complete memory enqueues in a narrow window, but anything added after
+        # the drain copies the buffer only resets the debounce Timer
+        # (best-effort, same as today).
+        #
+        # No host-level pending/processing guard: ``shutdown_flush``
+        # short-circuits on a truly idle buffer (returns True immediately), so
+        # calling it unconditionally is cheap and keeps the in-flight-worker
+        # race entirely inside the backend (where the buffer lives) -- the host
+        # cannot "forget" that case the way a ``pending_count > 0``-only guard
+        # would (review #6 on the original PR).
+        #
+        # K8s caveat: ``shutdown_flush_timeout_seconds`` must fit inside the
+        # pod's ``terminationGracePeriodSeconds`` (channel stop + this drain +
+        # buffer), set on the gateway Helm deployment -- or K8s SIGKILLs the
+        # drain mid-flight and the loss this is fixing is silently re-introduced.
+        try:
+            app_cfg = get_app_config()
+            if app_cfg.memory.enabled:
+                from deerflow.agents.memory import get_memory_manager
+
+                manager = get_memory_manager()
+                flush_timeout = app_cfg.memory.shutdown_flush_timeout_seconds
+                completed = await asyncio.to_thread(manager.shutdown_flush, flush_timeout)
+                if completed:
+                    logger.info("Memory queue flush completed within %.1fs", flush_timeout)
+                else:
+                    logger.warning(
+                        "Memory queue flush did not finish within %.1fs; remaining updates may be lost",
+                        flush_timeout,
+                    )
+        except Exception:
+            logger.exception("Failed to flush memory queue on shutdown")
+
     logger.info("Shutting down API Gateway")
 
 

backend/app/gateway/routers/memory.py

@@ -130,6 +130,7 @@ class MemoryConfigResponse(BaseModel):
     enabled: bool = Field(..., description="Whether the memory mechanism is enabled (call-site gate).")
     mode: Literal["middleware", "tool"] = Field(..., description="Memory operation mode: 'middleware' (passive per-turn LLM summarization) or 'tool' (model calls memory tools directly). Mechanism-level, applies to any backend.")
     injection_enabled: bool = Field(..., description="Whether memory is injected into the system prompt (call-site gate).")
+    shutdown_flush_timeout_seconds: float = Field(..., description="Hard budget (s) to drain pending memory updates on Gateway graceful shutdown; must fit inside the pod's K8s terminationGracePeriodSeconds.")
     manager_class: str = Field(..., description="Active memory backend selector (backend name or dotted path).")
     backend_config: dict = Field(..., description="Backend-private config (self-interpreted by the backend).")
 
@@ -362,6 +363,7 @@ async def get_memory_config_endpoint() -> MemoryConfigResponse:
         {
             "enabled": true,
             "injection_enabled": true,
+            "shutdown_flush_timeout_seconds": 30.0,
             "mode": "middleware",
             "manager_class": "deermem",
             "backend_config": {
@@ -380,6 +382,7 @@ async def get_memory_config_endpoint() -> MemoryConfigResponse:
         enabled=config.enabled,
         mode=config.mode,
         injection_enabled=config.injection_enabled,
+        shutdown_flush_timeout_seconds=config.shutdown_flush_timeout_seconds,
         manager_class=config.manager_class,
         backend_config=config.backend_config,
     )
@@ -406,6 +409,7 @@ async def get_memory_status(http_request: Request) -> MemoryStatusResponse:
             enabled=config.enabled,
             mode=config.mode,
             injection_enabled=config.injection_enabled,
+            shutdown_flush_timeout_seconds=config.shutdown_flush_timeout_seconds,
             manager_class=config.manager_class,
             backend_config=config.backend_config,
         ),

backend/packages/harness/deerflow/agents/memory/backends/deermem/deer_mem.py

@@ -237,6 +237,19 @@ class DeerMem(MemoryManager):
         """Not implemented this phase (no distinct export yet; /export routes via get_memory)."""
         raise NotImplementedError("DeerMem.export_memory is not implemented yet")
 
+    # ── Lifecycle ───────────────────────────────────────────────────────
+    def shutdown_flush(self, timeout: float) -> bool:
+        """Drain the debounce queue within ``timeout`` on graceful shutdown.
+
+        Delegates to the queue's bounded synchronous flush, which joins an
+        in-flight worker first (so contexts a debounce Timer already pulled out
+        of the queue are not lost on exit) and otherwise drains the queue on a
+        daemon thread with a real hard timeout (the memory-update LLM call is
+        synchronous and cannot be interrupted). Returns ``True`` only when the
+        drain genuinely finished within ``timeout``.
+        """
+        return self._queue.flush_sync(timeout)
+
     # ── DeerMem-internal (NOT on the ABC; reached via hasattr probing) ───
     def warm(self) -> bool:
         """Pre-warm DeerMem-specific resources (the tiktoken encoding cache).

backend/packages/harness/deerflow/agents/memory/backends/deermem/deermem/core/queue.py

@@ -47,6 +47,11 @@ class MemoryUpdateQueue:
         self._lock = threading.Lock()
         self._timer: threading.Timer | None = None
         self._processing = False
+        # Thread currently running ``_process_queue`` (None when idle). ``flush_sync``
+        # joins an in-flight worker instead of reporting a false-positive "completed"
+        # while contexts it already pulled out of the queue are still being processed
+        # (and would be lost on exit). See ``flush_sync`` step (1).
+        self._processing_thread: threading.Thread | None = None
         self._reprocess_pending = False
 
     @staticmethod
@@ -172,8 +177,15 @@ class MemoryUpdateQueue:
         self._timer.daemon = True
         self._timer.start()
 
-    def _process_queue(self) -> None:
+    def _process_queue(self, *, skip_inter_item_delay: bool = False) -> None:
-        """Process all queued conversation contexts."""
+        """Process all queued conversation contexts.
+
+        Args:
+            skip_inter_item_delay: When set, skip the inter-item rate-limit
+                ``time.sleep``. Intended for the shutdown-drain path
+                (:meth:`flush_sync`), which races a bounded timeout and should
+                not waste budget sleeping between items.
+        """
         with self._lock:
             if self._processing:
                 # Another worker is already draining the queue. Instead of
@@ -188,12 +200,15 @@ class MemoryUpdateQueue:
                 return
 
             self._processing = True
+            self._processing_thread = threading.current_thread()
             contexts_to_process = self._queue.copy()
             self._queue.clear()
             self._timer = None
 
         logger.info("Processing %d queued memory updates", len(contexts_to_process))
 
+        succeeded = 0
+        failed = 0
         try:
             for context in contexts_to_process:
                 try:
@@ -208,35 +223,122 @@ class MemoryUpdateQueue:
                         trace_id=context.trace_id,
                     )
                     if success:
+                        succeeded += 1
                         logger.info("Memory updated successfully for thread %s (trace_id=%s)", context.thread_id, context.trace_id)
                     else:
+                        failed += 1
                         logger.warning("Memory update skipped/failed for thread %s (trace_id=%s)", context.thread_id, context.trace_id)
                 except Exception as e:
+                    failed += 1
                     logger.error("Error updating memory for thread %s (trace_id=%s): %s", context.thread_id, context.trace_id, e)
 
-                # Small delay between updates to avoid rate limiting
+                # Small delay between updates to avoid rate limiting.
-                if len(contexts_to_process) > 1:
+                # Skipped on the shutdown-drain path, which races a bounded
+                # timeout and should spend that budget on LLM calls, not on
+                # sleeping between items.
+                if not skip_inter_item_delay and len(contexts_to_process) > 1:
                     time.sleep(0.5)
-
         finally:
+            # Summary count disambiguates "drained" (queue emptied) from "saved"
+            # (every extraction persisted): per-item ``update_memory`` failures are
+            # swallowed above, so without this an operator debugging missing
+            # memories would see only the happy-path "Processing N" line.
+            if succeeded or failed:
+                logger.info("Memory update batch done: %d succeeded, %d failed", succeeded, failed)
             with self._lock:
                 self._processing = False
+                self._processing_thread = None
                 if self._reprocess_pending:
                     self._reprocess_pending = False
                     if self._queue:
                         self._schedule_timer(0)
 
-    def flush(self) -> None:
+    def flush(self, *, skip_inter_item_delay: bool = False) -> None:
         """Force immediate processing of the queue.
 
         This is useful for testing or graceful shutdown.
+
+        Args:
+            skip_inter_item_delay: Forwarded to :meth:`_process_queue`; skip the
+                inter-item rate-limit sleep. Intended for the shutdown-drain
+                path (:meth:`flush_sync`).
         """
         with self._lock:
             if self._timer is not None:
                 self._timer.cancel()
                 self._timer = None
 
-        self._process_queue()
+        self._process_queue(skip_inter_item_delay=skip_inter_item_delay)
+
+    def flush_sync(self, timeout: float) -> bool:
+        """Best-effort synchronous flush bounded by ``timeout`` seconds.
+
+        Unlike :meth:`flush_nowait` (which only schedules a daemon timer that
+        is killed on process exit), this runs :meth:`flush` on a daemon thread
+        and waits up to ``timeout`` seconds for it to finish. Intended for
+        graceful shutdown: without it, any updates enqueued since the last
+        timer fire are lost on restart / rolling deploy / SIGTERM, because the
+        queue is pure in-memory and the debounce Timer is a daemon thread.
+
+        The drain accounts for two races a naive ``flush()`` would miss:
+
+        - **In-flight worker.** If the debounce Timer already fired, an
+          ``_process_queue`` worker is mid-LLM-call holding contexts it already
+          pulled out of the queue (``_processing=True``, queue empty). ``flush``
+          alone would see ``_processing=True``, no-op, and report success while
+          that worker is still running and likely killed on exit. So we join
+          the in-flight worker first (bounded by the remaining budget).
+        - **Failed flush.** ``flush`` makes a synchronous LLM call that can
+          raise; success is tracked on the happy path only, so the return value
+          matches the docstring's "completed".
+
+        Note: steps (1) and (3) share the same ``deadline`` budget. A slow
+        in-flight worker can consume most/all of it, leaving step (3) to no-op;
+        ``timeout`` must therefore cover both a slow in-flight worker *and* the
+        remaining queue (best-effort: any tail not drained in budget is dropped,
+        same failure direction as no flush, scoped to the tail).
+
+        Returns ``True`` only if the drain genuinely finished (queue empty, no
+        worker still running, flush did not raise) within ``timeout``.
+        """
+        deadline = time.monotonic() + timeout
+
+        # (1) Wait for an in-flight _process_queue first (bounded). Otherwise
+        # flush() would see _processing=True, no-op, and we would report
+        # success while that worker is still mid-LLM-call on a daemon thread
+        # that exit will kill — losing the contexts it already pulled out.
+        with self._lock:
+            in_flight = self._processing_thread
+        if in_flight is not None:
+            in_flight.join(timeout=max(0.0, deadline - time.monotonic()))
+
+        # (2) Genuine idle: nothing pending and no worker still running.
+        if self.pending_count == 0 and not self.is_processing:
+            return True
+
+        # (3) Drain the queue on a daemon thread so the timeout is a real hard
+        # stop: flush() makes a synchronous LLM call that cannot be
+        # interrupted, so we wait on Event.wait, not on Thread.join.
+        success = False
+        done = threading.Event()
+
+        def _run() -> None:
+            nonlocal success
+            try:
+                self.flush(skip_inter_item_delay=True)
+                success = True
+            except Exception:
+                logger.exception("Memory queue flush failed during shutdown drain")
+            finally:
+                done.set()
+
+        worker = threading.Thread(target=_run, name="memory-shutdown-flush", daemon=True)
+        worker.start()
+        finished = done.wait(timeout=max(0.0, deadline - time.monotonic()))
+        if not finished:
+            return False
+        # flush() returned; only report success if no worker raced back in.
+        return bool(success) and not self.is_processing
 
     def flush_nowait(self) -> None:
         """Start queue processing immediately in a background thread."""
@@ -256,6 +358,7 @@ class MemoryUpdateQueue:
                 self._timer = None
             self._queue.clear()
             self._processing = False
+            self._processing_thread = None
             self._reprocess_pending = False
 
     @property

backend/packages/harness/deerflow/agents/memory/backends/noop/noop_manager.py

@@ -157,6 +157,11 @@ class NoopMemoryManager(MemoryManager):
     ) -> dict[str, Any]:
         return _empty_memory()
 
+    # ── Lifecycle ───────────────────────────────────────────────────────
+    def shutdown_flush(self, timeout: float) -> bool:
+        """Nothing is ever queued, so shutdown drain is a clean no-op success."""
+        return True
+
     # ── Optional DeerMem-internal capabilities (NOT on the ABC) ──────────
     # The host gateway discovers these via ``hasattr(manager, "<name>")`` and
     # returns 501 when absent. Implement the ones your backend supports so the

backend/packages/harness/deerflow/agents/memory/manager.py

@@ -188,6 +188,34 @@ class MemoryManager(ABC):
     ) -> dict[str, Any]:
         """Export the memory document for the bucket. *stub* this phase (no caller yet)."""
 
+    # ── Lifecycle ───────────────────────────────────────────────────────
+    @abstractmethod
+    def shutdown_flush(self, timeout: float) -> bool:
+        """Best-effort bounded drain of pending updates on graceful shutdown.
+
+        Runs on the Gateway shutdown path (after IM channels and the scheduler
+        stop, so no new IM/scheduler updates arrive during the drain) to flush
+        updates still sitting in the backend's debounce buffer. Without it, any
+        update enqueued since the last timer fire is lost on restart / rolling
+        deploy / SIGTERM, because the buffer is pure in-memory and the debounce
+        worker is a daemon thread killed on process exit.
+
+        Implementations must honour a *hard* ``timeout``: the drain makes a
+        synchronous LLM call that cannot be interrupted, so the caller (the
+        Gateway lifespan) needs a real upper bound that lines up with the K8s
+        ``terminationGracePeriodSeconds`` (the drain must finish inside the pod
+        grace window, or K8s SIGKILLs it mid-drain and the loss the drain is
+        fixing is silently re-introduced).
+
+        Returns ``True`` if the drain genuinely finished within ``timeout``
+        (buffer empty, no worker still running, no exception); ``False`` on
+        timeout or failure (the caller logs a warning and proceeds to exit --
+        any unfinished tail is dropped, strictly better than no flush). A
+        backend with no pending work (or no buffer at all) returns ``True``
+        immediately, so the host may call this unconditionally when memory is
+        enabled without gating on backend-private queue state.
+        """
+
 
 # ── Backend discovery (drop-in) ───────────────────────────────────────────
 def _scan_backends() -> dict[str, type[MemoryManager]]:

backend/packages/harness/deerflow/client.py

@@ -1342,6 +1342,7 @@ class DeerFlowClient:
             "enabled": config.enabled,
             "mode": config.mode,
             "injection_enabled": config.injection_enabled,
+            "shutdown_flush_timeout_seconds": config.shutdown_flush_timeout_seconds,
             "manager_class": config.manager_class,
             "backend_config": config.backend_config,
         }

backend/packages/harness/deerflow/config/memory_config.py

@@ -4,7 +4,8 @@ DeerMem-private fields live in ``backends/deermem/config.py`` (``DeerMemConfig``
 reached via ``backend_config`` (a dict the factory passes to the backend's
 ``__init__``). This module holds ONLY the host-shared fields every backend /
 call site / factory reads: ``enabled`` / ``injection_enabled`` /
-``manager_class`` / ``backend_config``. Keeping the shared schema slim is what
+``shutdown_flush_timeout_seconds`` / ``manager_class`` / ``backend_config``.
+Keeping the shared schema slim is what
 makes backends swappable and portable (DeerMem's knobs do not leak onto the
 shared contract).
 """
@@ -17,7 +18,7 @@ from pydantic import BaseModel, Field
 logger = logging.getLogger(__name__)
 
 # Host-shared MemoryConfig fields (read by every backend / call site / factory).
-_SHARED_FIELDS = frozenset({"enabled", "mode", "injection_enabled", "manager_class", "backend_config"})
+_SHARED_FIELDS = frozenset({"enabled", "mode", "injection_enabled", "shutdown_flush_timeout_seconds", "manager_class", "backend_config"})
 
 # DeerMem-private fields that used to live at the top level of `memory:` in
 # config.yaml (pre-abstraction). On load they are auto-migrated into
@@ -66,6 +67,23 @@ class MemoryConfig(BaseModel):
         default=True,
         description="Whether to inject memory into the system prompt (call-site gate).",
     )
+    shutdown_flush_timeout_seconds: float = Field(
+        default=30.0,
+        ge=1.0,
+        le=300.0,
+        description=(
+            "Hard time budget (seconds) for draining the memory backend's "
+            "pending-update buffer during Gateway graceful shutdown. The drain "
+            "makes one LLM call per pending item, so large IM batches may need "
+            "a higher value. Must fit inside the pod's K8s "
+            "terminationGracePeriodSeconds (together with channel/scheduler "
+            "stop) or K8s SIGKILLs the drain mid-flight. The drain runs on a "
+            "daemon thread, so on timeout the process proceeds to exit and any "
+            "unfinished tail is dropped (same failure direction as no flush, "
+            "scoped to the tail). Host-shared (not backend-private): the host "
+            "owns the lifespan budget and the K8s grace relationship."
+        ),
+    )
     manager_class: str = Field(
         default="deermem",
         description=(

backend/tests/test_gateway_lifespan_shutdown.py

@@ -10,6 +10,7 @@ signal-reentrancy deadlock described in
 from __future__ import annotations
 
 import asyncio
+import logging
 from contextlib import asynccontextmanager
 from types import SimpleNamespace
 from unittest.mock import AsyncMock, MagicMock, patch
@@ -35,6 +36,9 @@ async def _run_lifespan_with_hanging_stop() -> float:
     app = FastAPI()
     startup_config = MagicMock()
     startup_config.log_level = "INFO"
+    # Keep this test focused on the channel-hang timing: skip the memory drain.
+    startup_config.memory.enabled = False
+    startup_config.memory.shutdown_flush_timeout_seconds = 5.0
     fake_service = MagicMock()
     fake_service.get_status = MagicMock(return_value={})
 
@@ -79,7 +83,7 @@ async def _run_lifespan_with_upload_staging_cleanup():
     from app.gateway.app import lifespan
 
     app = FastAPI()
-    startup_config = SimpleNamespace(log_level="INFO", memory=SimpleNamespace(token_counting="char"))
+    startup_config = SimpleNamespace(log_level="INFO", memory=SimpleNamespace(token_counting="char", enabled=False, shutdown_flush_timeout_seconds=30.0))
     fake_service = MagicMock()
     fake_service.get_status = MagicMock(return_value={})
     cleanup_upload_staging_files = MagicMock(return_value=2)
@@ -110,3 +114,78 @@ def test_lifespan_sweeps_upload_staging_files_on_startup():
     cleanup_upload_staging_files.assert_called_once_with()
     close_oidc_service.assert_awaited_once()
     stop_channel_service.assert_awaited_once()
+
+
+async def _run_lifespan_with_memory_flush(*, enabled: bool, flush_return: bool) -> MagicMock:
+    """Drive lifespan with a spied memory manager.shutdown_flush.
+
+    Returns the manager mock so the caller can assert the shutdown flush was
+    reached (and with what timeout). The host calls ``shutdown_flush``
+    unconditionally when memory is enabled -- there is no host-level
+    ``pending_count/is_processing`` gate, because the backend short-circuits on
+    an idle buffer and keeping the in-flight race inside the backend means the
+    host cannot "forget" it (review #6 on the original PR).
+    """
+    from app.gateway.app import lifespan
+
+    app = FastAPI()
+    startup_config = SimpleNamespace(
+        log_level="INFO",
+        memory=SimpleNamespace(
+            token_counting="char",
+            enabled=enabled,
+            shutdown_flush_timeout_seconds=5.0,
+        ),
+    )
+    fake_service = MagicMock()
+    fake_service.get_status = MagicMock(return_value={})
+    close_oidc_service = AsyncMock()
+    stop_channel_service = AsyncMock()
+
+    async def fake_start(_startup_config):
+        return fake_service
+
+    manager = MagicMock()
+    manager.shutdown_flush.return_value = flush_return
+
+    with (
+        patch("app.gateway.app.get_app_config", return_value=startup_config),
+        patch("app.gateway.app.get_gateway_config", return_value=MagicMock(host="x", port=0)),
+        patch("app.gateway.app.langgraph_runtime", _noop_langgraph_runtime),
+        patch("app.gateway.app.auth.close_oidc_service", close_oidc_service),
+        patch("app.channels.service.start_channel_service", side_effect=fake_start),
+        patch("app.channels.service.stop_channel_service", stop_channel_service),
+        patch("deerflow.agents.memory.get_memory_manager", return_value=manager),
+    ):
+        async with lifespan(app):
+            pass
+
+    return manager
+
+
+def test_lifespan_drains_memory_on_shutdown_with_configured_timeout(caplog) -> None:
+    """When memory is enabled, shutdown calls manager.shutdown_flush with the
+    configured timeout (asserts the timeout is forwarded, review #3) and logs
+    'completed' at INFO when the drain finishes."""
+    caplog.set_level(logging.INFO, logger="app.gateway.app")
+    manager = asyncio.run(_run_lifespan_with_memory_flush(enabled=True, flush_return=True))
+    manager.shutdown_flush.assert_called_once_with(5.0)
+    assert any(r.levelno == logging.INFO and "flush completed" in r.message for r in caplog.records)
+
+
+def test_lifespan_warns_when_memory_flush_does_not_finish(caplog) -> None:
+    """A False return (timeout/failure) is the path operators actually see when
+    K8s SIGKILLs the drain; the host must log a WARNING (not 'completed'), so
+    the loss risk is visible (review #3 False-branch coverage; review #2/#4
+    failed-flush semantics)."""
+    caplog.set_level(logging.WARNING, logger="app.gateway.app")
+    manager = asyncio.run(_run_lifespan_with_memory_flush(enabled=True, flush_return=False))
+    manager.shutdown_flush.assert_called_once_with(5.0)
+    assert any(r.levelno == logging.WARNING and "did not finish" in r.message for r in caplog.records)
+    assert not any("flush completed" in r.message for r in caplog.records)
+
+
+def test_lifespan_skips_memory_flush_when_disabled() -> None:
+    """memory.enabled=False skips the drain entirely."""
+    manager = asyncio.run(_run_lifespan_with_memory_flush(enabled=False, flush_return=True))
+    manager.shutdown_flush.assert_not_called()

backend/tests/test_memory_manager_pluggable.py

@@ -15,6 +15,7 @@ are independent of order.
 from __future__ import annotations
 
 from pathlib import Path
+from unittest.mock import MagicMock, patch
 
 import pytest
 
@@ -146,3 +147,47 @@ def test_empty_storage_path_factory_injects_runtime_home(tmp_path, monkeypatch)
     # per-user dir created under the injected runtime_home root
     user_dirs = [p.name for p in (tmp_path / "users").iterdir() if p.is_dir()]
     assert len(user_dirs) == 1
+
+
+def test_shutdown_flush_is_on_abc_and_noop_is_noop_success() -> None:
+    """``shutdown_flush`` is part of the MemoryManager ABC (every backend must
+    implement a bounded graceful-shutdown drain); noop has no buffer, so it
+    returns True immediately."""
+    assert "shutdown_flush" in MemoryManager.__abstractmethods__
+    reset_memory_manager()
+    set_memory_config(MemoryConfig(manager_class="noop"))
+    noop = get_memory_manager()
+    assert noop.shutdown_flush(1.0) is True
+    reset_memory_manager()
+
+
+def test_deermem_shutdown_flush_delegates_to_queue_flush_sync() -> None:
+    """DeerMem.shutdown_flush delegates to its queue's bounded flush_sync,
+    forwarding the host-owned timeout budget unchanged."""
+    reset_memory_manager()
+    set_memory_config(MemoryConfig(manager_class="deermem"))
+    deermem = get_memory_manager()
+    with patch.object(deermem._queue, "flush_sync", return_value=True) as spy:
+        assert deermem.shutdown_flush(7.0) is True
+        spy.assert_called_once_with(7.0)
+    reset_memory_manager()
+
+
+def test_deermem_shutdown_flush_drains_a_pending_update() -> None:
+    """End-to-end: a pending update sitting in DeerMem's queue is drained
+    within the timeout on shutdown_flush (the loss-on-exit bug the ABC method
+    fixes). The drain skips inter-item sleep so the budget goes to LLM calls."""
+    from deerflow.agents.memory.backends.deermem.deermem.core.queue import ConversationContext
+
+    reset_memory_manager()
+    set_memory_config(MemoryConfig(manager_class="deermem"))
+    deermem = get_memory_manager()
+    # Inject a mock updater so no real LLM call is made; both items "succeed".
+    mock_updater = MagicMock()
+    mock_updater.update_memory.return_value = True
+    deermem._queue._updater = mock_updater
+    deermem._queue._queue = [ConversationContext(thread_id=f"t{i}", messages=["m"], agent_name="lead_agent") for i in range(3)]
+    assert deermem.shutdown_flush(5.0) is True
+    assert deermem._queue.pending_count == 0
+    assert mock_updater.update_memory.call_count == 3
+    reset_memory_manager()

backend/tests/test_memory_queue.py

@@ -247,3 +247,159 @@ def test_process_queue_forwards_trace_id_to_updater() -> None:
         user_id=None,
         trace_id="trace-memory-1",
     )
+
+
+# ---------------------------------------------------------------------------
+# shutdown_flush / flush_sync (graceful-shutdown drain) — review carry-overs.
+# The queue is a daemon-timer + in-memory buffer, so anything pending at
+# process exit is lost. flush_sync drains it within a hard timeout, joining an
+# in-flight worker first so contexts a debounce Timer already pulled out of the
+# queue are not lost either.
+# ---------------------------------------------------------------------------
+
+_QUEUE_MODULE = "deerflow.agents.memory.backends.deermem.deermem.core.queue"
+
+
+def test_flush_sync_noop_on_empty_queue() -> None:
+    """flush_sync short-circuits and returns True when there is nothing to drain."""
+    queue = _queue()
+    assert queue.pending_count == 0
+    assert queue.flush_sync(timeout=5.0) is True
+
+
+def test_flush_sync_drains_pending_queue_and_returns_true() -> None:
+    """flush_sync runs the synchronous flush() and waits for it to finish."""
+    mock_updater = MagicMock()
+    mock_updater.update_memory.return_value = True
+    queue = _queue(mock_updater)
+    queue._queue = [ConversationContext(thread_id="thread-1", messages=["conversation"], agent_name="lead_agent")]
+
+    with (
+        patch(_QUEUE_MODULE + ".MemoryUpdater", create=True),
+        patch(_QUEUE_MODULE + ".time.sleep"),
+    ):
+        completed = queue.flush_sync(timeout=5.0)
+
+    assert completed is True
+    assert queue.pending_count == 0
+    mock_updater.update_memory.assert_called_once_with(
+        messages=["conversation"],
+        thread_id="thread-1",
+        agent_name="lead_agent",
+        correction_detected=False,
+        reinforcement_detected=False,
+        user_id=None,
+        trace_id=None,
+    )
+
+
+def test_flush_sync_returns_false_when_flush_exceeds_timeout() -> None:
+    """flush_sync does not block past ``timeout``; a slow flush returns False."""
+    queue = _queue()
+    queue._queue = [ConversationContext(thread_id="thread-1", messages=["conversation"], agent_name="lead_agent")]
+    release = threading.Event()
+
+    def _slow_flush() -> None:
+        # Block until the test releases us (well past the flush_sync timeout).
+        release.wait(timeout=5.0)
+
+    with patch.object(queue, "flush", side_effect=_slow_flush):
+        completed = queue.flush_sync(timeout=0.1)
+
+    assert completed is False
+    # The queue was not drained because flush() never returned.
+    assert queue.pending_count == 1
+    # Release the daemon thread so it does not linger past the test.
+    release.set()
+
+
+def _run_inflight_worker(queue: MemoryUpdateQueue, release: threading.Event) -> threading.Thread:
+    """Start a thread that mimics _process_queue's "pulled contexts, mid-LLM" state.
+
+    It claims ``_processing`` / ``_processing_thread`` (so the queue looks idle
+    by ``pending_count`` but a worker is in flight), blocks on ``release``,
+    then clears the flags on the way out.
+    """
+
+    def _inflight() -> None:
+        with queue._lock:
+            queue._processing = True
+            queue._processing_thread = threading.current_thread()
+        release.wait(timeout=5.0)
+        with queue._lock:
+            queue._processing = False
+            queue._processing_thread = None
+
+    thread = threading.Thread(target=_inflight, name="fake-inflight-worker", daemon=True)
+    thread.start()
+    # Wait until the fake worker has claimed _processing.
+    while not queue.is_processing:
+        time.sleep(0.005)
+    return thread
+
+
+def test_flush_sync_waits_for_inflight_worker_and_returns_false_if_unfinished() -> None:
+    """flush_sync must not report success while an in-flight _process_queue is
+    still mid-LLM-call — the contexts it already pulled out would be lost on
+    exit. It joins the in-flight worker (bounded) and returns False when the
+    worker does not finish within the budget (review comment #1)."""
+    queue = _queue()
+    release = threading.Event()
+    inflight = _run_inflight_worker(queue, release)
+
+    try:
+        completed = queue.flush_sync(timeout=0.2)
+    finally:
+        release.set()
+        inflight.join(timeout=5.0)
+
+    assert completed is False
+
+
+def test_flush_sync_returns_true_when_inflight_worker_finishes_in_budget() -> None:
+    """When the in-flight worker finishes within the budget, flush_sync joins it
+    and reports success (review comment #1, positive case)."""
+    queue = _queue()
+    release = threading.Event()
+    inflight = _run_inflight_worker(queue, release)
+
+    # Let the in-flight worker finish well within the budget.
+    release.set()
+
+    completed = queue.flush_sync(timeout=5.0)
+    inflight.join(timeout=5.0)
+
+    assert completed is True
+    assert queue.is_processing is False
+    assert queue._processing_thread is None
+
+
+def test_flush_sync_returns_false_when_flush_raises() -> None:
+    """flush_sync reports failure (not success) when flush() raises, so the
+    caller never logs a contradictory 'completed' next to the exception
+    (review comment #2)."""
+    queue = _queue()
+    queue._queue = [ConversationContext(thread_id="thread-1", messages=["conversation"], agent_name="lead_agent")]
+
+    with patch.object(queue, "flush", side_effect=RuntimeError("boom")):
+        completed = queue.flush_sync(timeout=5.0)
+
+    assert completed is False
+
+
+def test_flush_sync_skips_inter_item_delay_on_drain_path() -> None:
+    """On the shutdown-drain path the per-item rate-limit sleep is skipped so
+    the bounded timeout covers as many items as possible (review comment #5)."""
+    mock_updater = MagicMock()
+    mock_updater.update_memory.return_value = True
+    queue = _queue(mock_updater)
+    queue._queue = [ConversationContext(thread_id=f"thread-{i}", messages=["conversation"], agent_name="lead_agent") for i in range(3)]
+
+    with patch(_QUEUE_MODULE + ".time.sleep") as mock_sleep:
+        completed = queue.flush_sync(timeout=5.0)
+
+    assert completed is True
+    assert queue.pending_count == 0
+    # No inter-item rate-limit sleep on the drain path.
+    mock_sleep.assert_not_called()
+    assert mock_updater.update_memory.call_count == 3

config.example.yaml

@@ -15,7 +15,7 @@
 # ============================================================================
 # Bump this number when the config schema changes.
 # Run `make config-upgrade` to merge new fields into your local config.yaml.
-config_version: 25
+config_version: 26
 
 # ============================================================================
 # Logging
@@ -1453,6 +1453,7 @@ summarization:
 # Shared fields (host level, backend-agnostic):
 #   enabled             - Master switch for the memory mechanism (call-site gate)
 #   injection_enabled   - Whether to inject memory into the system prompt (call-site gate)
+#   shutdown_flush_timeout_seconds - Hard budget (s) to drain pending updates on Gateway graceful shutdown (default: 30)
 #   manager_class       - Backend selector: registered name (deermem/noop) or dotted path
 #   backend_config      - Backend-private config dict (passthrough; each backend self-interprets)
 #
@@ -1476,6 +1477,12 @@ summarization:
 memory:
   enabled: true
   injection_enabled: true # Whether to inject memory into system prompt
+  # Hard budget (seconds) to drain pending memory updates on Gateway graceful
+  # shutdown. Each pending item does one LLM call, so large IM batches may need
+  # more. Must fit inside the pod's K8s terminationGracePeriodSeconds (channel
+  # stop + this drain + buffer) or K8s SIGKILLs the drain -- set that on the
+  # gateway Helm deployment (see deploy/helm/deer-flow). Default 30s.
+  shutdown_flush_timeout_seconds: 30.0
   # Memory backend selector. Either a registered backend name (matching a
   # backends/<name>/ folder that exposes MANAGER_CLASS, e.g. deermem / noop)
   # or a dotted import path to a MemoryManager subclass.

deploy/helm/deer-flow/README.md

@@ -124,7 +124,7 @@ they resolve from the `secrets` map):
 
 ```yaml
 config: |
-  config_version: 25
+  config_version: 26
   models:
     - name: gpt-4
       use: langchain_openai:ChatOpenAI
@@ -227,6 +227,7 @@ kubectl -n deer-flow exec deploy/deer-flow-provisioner -- curl -s localhost:8002
       username: deerflow
       password: changeme
   ```
+- **Graceful shutdown & memory drain.** The gateway pod sets `terminationGracePeriodSeconds` (default 45s, overridable via `gateway.terminationGracePeriodSeconds`) plus an optional `preStop` sleep (`gateway.preStopSleepSeconds`, default 5s). The grace period MUST exceed the Gateway's graceful-shutdown work — channel stop (~5s) plus the memory-queue drain (`memory.shutdown_flush_timeout_seconds`, default 30s) plus a buffer — because the drain runs on a daemon thread and K8s SIGKILLs anything still running at the end of the grace window. K8s defaults to 30s, which SIGKILLs the drain mid-flight and silently re-introduces the memory loss the drain is fixing. **When you raise `memory.shutdown_flush_timeout_seconds`, raise `gateway.terminationGracePeriodSeconds` to match** (channel stop + drain + buffer).
 - **Gateway replicas.** Postgres + the Redis stream bridge together make the
   gateway's *persisted* state (checkpointer + run/thread metadata) and *live
   stream* path cross-pod-safe. The default is still 1 replica: **do not raise

deploy/helm/deer-flow/templates/gateway-deployment.yaml

@@ -24,6 +24,10 @@ spec:
         checksum/extensions: {{ include "deer-flow.extensionsChecksum" . }}
     spec:
       {{- include "deer-flow.imagePullSecrets" . | nindent 6 }}
+      # MUST exceed channel stop (~5s) + memory shutdown_flush_timeout_seconds
+      # (default 30s) + buffer, or K8s SIGKILLs the memory drain mid-flight
+      # (values.yaml gateway.terminationGracePeriodSeconds). K8s defaults to 30s.
+      terminationGracePeriodSeconds: {{ .Values.gateway.terminationGracePeriodSeconds | default 45 }}
       securityContext:
         # Non-root: the gateway image has no USER directive and would otherwise
         # run as root. uid/gid 1000; fsGroup 1000 so the .deer-flow PVC is
@@ -145,6 +149,16 @@ spec:
               port: http
             initialDelaySeconds: 30
             periodSeconds: 20
+          {{- /* preStop sleep so the Service/ingress deregisters the pod and
+                  in-flight requests drain before SIGTERM begins graceful shutdown
+                  (which runs the memory queue drain). 0 / unset disables it.
+                  Counts against terminationGracePeriodSeconds. */ -}}
+          {{- if and .Values.gateway.preStopSleepSeconds (gt (.Values.gateway.preStopSleepSeconds | int) 0) }}
+          lifecycle:
+            preStop:
+              exec:
+                command: ["sleep", "{{ .Values.gateway.preStopSleepSeconds | int }}"]
+          {{- end }}
           {{- with .Values.gateway.resources }}
           resources:
             {{- toYaml . | nindent 12 }}

deploy/helm/deer-flow/values.yaml

@@ -28,6 +28,18 @@ gateway:
   replicas: 1               # Safe default. Postgres + the Redis stream bridge are
                             # wired, but multi-replica needs issue #3948's run-control
                             # work (cancel/dedup/reconcile) — see README "Gateway replicas".
+  # Pod termination grace period (seconds). MUST exceed the Gateway's graceful
+  # shutdown work: channel stop (~5s) + memory queue drain
+  # (memory.shutdown_flush_timeout_seconds, default 30s) + buffer. K8s defaults
+  # to 30s, which SIGKILLs the memory drain mid-flight and silently re-introduces
+  # the memory loss that shutdown_flush is fixing. Raise this when you raise the
+  # memory timeout. See gateway-deployment.yaml + README "Graceful shutdown".
+  terminationGracePeriodSeconds: 45
+  # preStop hook sleep (seconds) before SIGTERM, so the Service/ingress
+  # deregisters the pod and in-flight HTTP requests drain before graceful
+  # shutdown begins (which runs the memory queue drain). 0 disables the hook.
+  # Counts against terminationGracePeriodSeconds.
+  preStopSleepSeconds: 5
   resources:
     requests:
       cpu: 200m
@@ -221,7 +233,7 @@ ingress:
 # -- DeerFlow config.yaml content. Secrets MUST stay as $VAR references — never
 #    inline literal secret values here. The default enables provisioner sandbox.
 config: |
-  config_version: 25
+  config_version: 26
   log_level: info
 
   models: []