2 Commits

Author	SHA1	Message	Date
Nan Gao	dc6f0d7421	fix(memory): persist the DeerMem extraction watermark across restarts and replicas (#7205) ... * fix(memory): persist the DeerMem extraction watermark across restarts and replicas The updater's per-thread extraction watermark lived only in an in-process LRU, so a restart or a second Gateway replica on the same storage root fed already-extracted turns to the memory LLM again. FileMemoryStorage now keeps it in users/{user}/agents/{agent}/.metadata/ watermarks.json, keyed by thread id, written atomically under the per-user lock without touching memory.json and capped by watermark_max_keys. Each update feeds after the furthest of the LRU entry and the persisted one, and writes through after an extraction or a pre-screen skip; bypass_watermark is unchanged and sidecar IO failures fall back to the LRU. Memory clears remove the file with the other .metadata sidecars. Content-based message identities are stored as SHA-256 digests so the sidecar holds no conversation text. * docs(changelog): link the memory watermark entry to #7204 and #7205 * fix(memory): never replace an unreadable watermark sidecar set_extraction_watermark rewrites the whole watermarks.json, and the shared sidecar reader turns an OSError into an empty mapping, so a failed read could replace every other thread's watermark with the current one (or delete the file). The watermark sidecar now has its own reader: a missing file reads as empty, malformed content reads as empty (nothing to keep; the next write repairs it), and any other OSError propagates before anything is written. The updater already logs that and keeps the extraction successful. * fix(memory): keep the persisted watermark when the last message has no identity _message_identity returns None for a last message with neither an id nor text. The updater wrote that None through, which deleted the thread's entry from watermarks.json, losing a watermark a peer may have advanced and forcing a full feed on every replica. The write-through now skips None: the persisted entry still marks an extracted prefix and the furthest-of-both read keeps using it.	2026-10-21 07:26:44 +08:00
lllyfff	8145d66a33	feat(memory): memory message processing (#4447) ... * feat(memory): signals-based update pipeline + always-on watermark/trivial filter Refactor the DeerMem memory update pipeline (message_processing -> queue -> updater) around a signals frozenset seam, replacing the (filtered, correction_detected, reinforcement_detected) 3-tuple with (filtered, signals: frozenset[str]) end to end. message_processing: - Externalize signal-detection patterns to YAML (message_patterns/*.yaml). - Extend signals from correction/reinforcement to a 6-class set (correction/reinforcement/preference/identity/goal/decision); detect_signals returns a frozenset aligned with the fact category enum. - Pure-acknowledgment turns ("ok"/"好的"/...) are always filtered out before enqueue (whole-message fullmatch), saving an extraction LLM call. queue (core/queue.py): - In-memory list + debounce timer, with flush_sync (graceful-shutdown drain that joins an in-flight worker under a hard timeout) and queue_max_depth backpressure (signal-bearing updates always admitted; QueueFull otherwise). - Same-key updates coalesce with a signal union; per-batch success/fail summary. updater (core/updater.py): - head500+tail500 message truncation (replaces the 1000-char head chop). - Always-on per-thread watermark: feed only messages added since the last extraction. The watermark is in-memory and is not advanced on failure, so a failed/lost update is re-fed on the next conversation turn. - [MANUAL] prompt marker for user-authored facts (source.type="manual"). - Post-invoke extraction_callback (host-injected) emitting facts_extracted / facts_accepted / rejected_low_confidence; the host default logs metrics and flags >60% rejection. Confidence filtering remains in _apply_updates (the existing fact_confidence_threshold check); there is no separate write gate. Consolidation stays opt-in (lossy). The ABC add/add_nowait signature is unchanged, so the summarization flush hook and host are unaffected. Tests: add test_message_processing_signals, test_updater_truncation, test_updater_watermark; update queue/updater/consolidation/staleness/pluggable tests for the signals seam. Co-Authored-By: Claude <noreply@anthropic.com> * fix(memory): harden update pipeline per PR review - Catch QueueFull in DeerMem.add/add_nowait so backpressure degrades to 'update skipped' instead of propagating into after_agent / summarization_hook and breaking the agent run (peer middlewares self-guard; MemoryMiddleware was the lone exception). Emergency (add_nowait) always admits under backpressure -- its data cannot be re-fed next turn. - Rewrite the watermark from index-based to content/identity-based (_message_identity + _feed_after_watermark) so it stays correct when summarization removes the conversation front -- an index watermark pointed at the wrong message and silently skipped un-extracted tail turns. The emergency flush bypasses the watermark (bypass_watermark on ConversationContext, threaded through update_memory) and coexists with (does not replace) a pending normal update, so a flush cannot drop a pending update's un-extracted tail. - Populate facts_accepted / rejected_low_confidence inside _apply_updates at the real confidence-filter site (passed_threshold) instead of re-deriving the threshold in _finalize_update -- eliminates metric drift. - Emit extraction metrics in a finally with an 'attempted' flag so exception failures (parse error, apply_changes raise after retry) are observable, not only the happy path. - Re-detect signals on the post-watermark feed for the extraction hint so it no longer references turns the LLM cannot see; admission-time signals still drive backpressure. - Move the post-batch reschedule inside the queue lock to close a non-atomic self._timer race with a concurrent add. Co-Authored-By: Claude <noreply@anthropic.com> * fix(memory): address follow-up review nits (LRU, metric name, docstring) - Bound the in-memory watermark cache with a configurable LRU (watermark_max_keys, default 4096, 0=unbounded). A dropped key re-extracts one batch on that thread's next turn (the documented restart behavior), so eviction is safe and preserves the content-identity watermark's front-removal guarantee. Adds _watermark_get/_watermark_set helpers and a bounded-LRU regression test. - Rename the extraction metric facts_accepted -> facts_passed_confidence so the name matches what the >60% rejection-rate warning assumes (a confidence-gate signal, not a persisted-fact count); drop the stale "historical semantics" justification. Brand-new callback, one consumer. - Fix the stale test_message_processing_signals module docstring: the signals seam is already swapped to frozenset, and a stale stage-numbering prefix is removed. Co-Authored-By: Claude <noreply@anthropic.com> --------- Co-authored-by: Claude <noreply@anthropic.com>	2026-07-26 21:16:36 +08:00