feat(memory): deferred task-conditioned injection (flag-gated) + three-leg contextual gate#84
Merged
Conversation
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…latch, all-tools matcher Wires the three-leg evidence-gated scorer (BM25/vector/keyword-fallback) into contextual_inject.py: builds a SyncEmbedder(down_state_file=...) when embeddings_backend=ollama and calls retrieve_relevant with the new conn/sync_embedder/vec_floor signature (no more min_hits). Adds SeenStore.pretool_fired()/mark_pretool_fired() so PreToolUse does real retrieval work only once per session; later PreToolUse events short-circuit via a module-local _SkipInjection sentinel before touching the DB. Widens install_hooks.py's PreToolUse matcher from "Skill|Task|Write" to None (unscoped) since the latch makes an all-tools matcher cheap. Also restores the blank-query short-circuit in retrieve_relevant that the Task 4 rewrite dropped (a contentless query has no evidence by definition), and fixes a SeenStore._load() bug found while testing the latch: it silently dropped the pretool_fired key on every reload because it only carried forward "turn"/"seen". Updates the 5 tests in test_contextual_inject.py that assumed the old keyword min-hits gate (BM25 now matches on token overlap, so test_below_floor_injects_nothing needed genuinely zero-evidence content), adds latch coverage, and fixes the golden hook-shape assertions in test_install_hooks.py (both) and test_setup_sh.py for the matcher change. Also updates test_agentcore_semantic_parity.py's stale min_hits= call, collateral from the Task 4 signature change but outside its declared scope. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
tests/hooks/test_contextual_inject.py's bm_home fixture never pinned BETTER_MEMORY_EMBEDDINGS_BACKEND, so every non-blank-query test relied on config.py's actual default (ollama): the hook built a real SyncEmbedder and issued an HTTP call to localhost:11434 on every run. This worked only because no test seeds embeddings, so a connection-refused/failed embed degrades silently to the qvec=None fallback path with no observable difference — a live Ollama on that port would have gone unnoticed too. Pins BETTER_MEMORY_EMBEDDINGS_BACKEND=sqlite in the fixture, mirroring tests/ui/conftest.py's pin, and corrects the stale comment on test_below_floor_injects_nothing that claimed sqlite was the config default rather than the fixture's pin. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…cy fixes Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…ktick detection Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Sync website/architecture.md, configuration.md, mcp-tools.md with the feat/deferred-injection branch: BETTER_MEMORY_INJECT_MODE (legacy default | deferred bootstrap), the three-leg evidence-gated contextual scorer (BM25 / vector cosine against BETTER_MEMORY_CONTEXT_VEC_FLOOR / keyword fallback), deprecate BETTER_MEMORY_CONTEXT_MIN_HITS, document the PreToolUse-widened-to-all-tools once-per-session latch, the file-persisted Ollama outage cooldown shared across hook processes, and via_exploration exposure tagging excluded from the headline usefulness metric.
…string, import hoist Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?s=60&v=4){kind=small}
There was a problem hiding this comment.
🔴 Claude BugBot Analysis
Found 1 potential bug in this PR.
medium: 1
Found one confirmed cross-backend logic bug: the new three-leg evidence-gated scorer in services/relevant.py computes a Wilson-score ranking prior assuming times_overlooked/times_ignored fields exist on reflection/semantic records, but the AgentCore backend never populates (or even exposes) those fields for reflections, silently inflating agentcore-backed reflection rankings.
…icts for the Wilson prior Addresses PR #84 review: missing keys made the contextual prior compute a falsely-perfect hit rate on the AgentCore backend. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
There was a problem hiding this comment.
🟢 Claude BugBot Analysis
No new bugs found. The previously-reported Wilson-score bug is fixed: AgentCore's reflection dict now carries a real times_overlooked count and semantic memories fall through to the SemanticMemory dataclass's times_ignored=0 default (both intentional, documented, and consistent with the protocol contract in storage/protocol.py).
No bugs were detected in this PR.
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
Implements the deferred-injection design (
docs/superpowers/specs/2026-07-23-deferred-injection-design.md); absorbs the planned PR-C (CLAUDE.md rewrite + drift sentinel).BETTER_MEMORY_INJECT_MODE—legacy(default; today's bootstrap, byte-identical, test-pinned) |deferred(SessionStart injects only general-scope semantic memories + one index line; ~2 exposures/session instead of 7). Flag misparse fails safe to legacy.BETTER_MEMORY_CONTEXT_VEC_FLOOR=0.55 + Wilson prior) behind an evidence gate — a memory injects only with BM25 or vec evidence; the prior ranks but can never qualify alone; silent when nothing qualifies. Keyword fallback covers semantics-without-vectors and agentcore mode.CONTEXT_MIN_HITSdeprecated.SeenStore._loaddropped unknown keys, so any persisted flag silently reset per hook process).SyncEmbedder(hooks are a fresh process per firing — an in-process breaker resets every prompt).via_exploration, migration 0015) — excluded from the headline usefulness metric; conversion rate becomes measurable.Validation
Rollout
Merges with legacy as the live default — bootstrap behaviour unchanged on merge day. Post-merge: env-flag A/B (identical code, only the flag differs per arm, 24 sessions/arm) gates the flip to
deferred; fail-twice = stay legacy.🤖 Generated with Claude Code