feat(memory): session-level memory observer actor (spike)#410
Merged
Conversation
Phase 1 of memory formation fixes based on analysis of three production sessions that showed zero memory proposals and zero recall matches. Recall fixes: - Allow evidence class in deterministic retrieval (was hardcoded to durable_fact only, making 10/28 stored memories invisible) - Add baseline score (1.0) in candidate selector so SQL-matched candidates aren't silently dropped by zero-score threshold - Switch to audience-primary recall: remove domain as hard filter, use audience+boundary as security gates, add domain affinity boost for ranking. Removes ShouldWidenAcrossDomains two-path design. Formation fixes: - Revise observation sidecar prompt: add agent-derived findings classification category, evidence example, soften conservative bias - Tighten ProjectStatementPattern to reject conversational fragments via IsConversationalFragment prefix check (blocks junk like "Well I was going to has You do some Netclaw work") Observability: - Log progressive_recall_exhausted when all candidates already injected - Add eval debugging guidance to CLAUDE.md (failures are almost always instrumentation, rarely the model)
Adds SessionMemoryObserverActor — a persistent child actor that watches the conversation stream and distills memories when the session goes idle. Replaces the per-turn observation sidecar which produced proposalCount=0 on every prompt because it lacked session-level context. Observer design: - Persistent actor (journals proposed anchors for skip list durability) - Subscribes to same stream as SessionLogActor (user messages, assistant text, tool call names, recalled memories, turn boundaries) - Owns its own ReceiveTimeout (90s default, configurable) - Runs sidecar LLM call with full transcript + skip list - Sends proposals to parent → gate → curation actor pipeline - Token usage emitted through standard EmitUsageOutput pipeline Session actor changes: - Creates observer alongside curation actor in RecoveryCompleted - Forwards SendUserMessage and SessionOutput to observer - Forwards ObserverSystemContext for recalled memories - Handles SessionDistillationCompleted with token tracking - Removes: ObserveTurnForMemory, ObserveMemoryAsync, BuildStrongAssertions, _sidecarMemoryObserver field, MemoryObservationCompleted/Failed handlers Known limitation: no passivation-triggered distillation yet. The observer's idle timer (90s) is shorter than passivation timeout (~5min), so memories form during normal pauses. Final distillation before death requires formalizing the session actor's state machine (separate work).
- Double the sidecar timeout for distillation (full transcript is larger than single-turn prompts) - Preserve _hasNewContent on distillation failure so the observer retries on the next idle cycle instead of silently losing the content - DistillationFinished now carries a success flag for this purpose Known: Qwen 3.5 27B times out even at 180s on distillation calls. The architecture works correctly (fires, retries, guards concurrency). A faster or API-hosted model is needed for practical use.
…ault Qwen 3.5 27B needs more time for full-transcript distillation. 2x was still timing out. 5x (450s with default 90s sidecar config) gives the local model enough room. Confirmed: session D0AC6CKBK5K/1774388427.214549 successfully distilled 2 proposals (litellm-supply-chain-attack, litellm-mitigation-steps), routed through gate (accepted=2) and curation (created=2).
ApplyInlineCurationBatchAsync was not writing boundary or audience columns, causing all observer-distilled memories to have NULL values. Since recall filters on COALESCE(boundary, 'boundary:legacy-restricted'), NULL-boundary documents were invisible to Slack sessions that expect 'boundary:trusted-instance'. Also fix HandleDistillationResult to resolve audience from session ID when _currentTurnSource is null (distillation fires after idle, turn context is cleared). Slack → Team, SignalR → Personal, etc.
5 tasks
This was referenced Mar 25, 2026
Merged
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Spike: adds
SessionMemoryObserverActor— a persistent child actor that watches theconversation stream and distills memories when the session goes idle. Replaces per-turn
observation which produced
proposalCount=0on every prompt.tool call names, recalled memories, turn boundaries)
Builds on PR #409 (Phase 1 recall fixes).
Known limitation
No passivation-triggered distillation. The observer's idle timer (90s) fires before
the session's passivation timeout (~5min), so memories form during normal pauses.
Final distillation before session death requires formalizing the session actor's state
machine — tracked in #TBD.
Test plan
dotnet slopwatch analyzecleannetclaw -pexercises → verifyproposalCount>0in daemon logs