feat: attention-first memory injection redesign#77
Merged
Conversation
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…confidence-scored) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…ts/max_items/reinject_turns)
…ounters Widens session_memory_exposure.source CHECK to admit 'contextual' and seeds 5 contextual_* rating_diagnostics counters, via the standard table-recreation pattern (SQLite can't ALTER a CHECK constraint). Numbered 0012 not 0011 (as the SDD task brief specified) because 0011_trigram_fts.sql landed on main first and claimed that version. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…hread os._exit(0)-killed the pytest process mid-suite Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…oth backends Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…-items cap Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…t-flip, rating affordance
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…l ids and age stamps
…p slimming Document the four new injection-tuning env vars (BOOTSTRAP_TOP_N, CONTEXT_MIN_HITS, CONTEXT_MAX_ITEMS, CONTEXT_REINJECT_TURNS), the contextual_inject hook (hits-x-activation scoring, seen-file dedup, project-memory block, contextual exposure source), bootstrap top-N slimming, and the per-source rating directive. All facts verified against the code landed in Tasks 1-11. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
… state/ layout REINJECT_TURNS docs implied a turn = one prompt-response cycle; a turn is actually one firing of the contextual_inject hook (each user prompt, plus each matched Skill|Task|Write tool call in mode both). Also note the contextual_* rating_diagnostics counters are per-firing not per-item, and add the state/ directory (context_seen_<session_id>.json) to the BETTER_MEMORY_HOME filesystem-layout tree in configuration.md.
…n agentcore - tests/hooks/test_contextual_inject.py: make the bm_home fixture autouse so every test in the module runs against a tmp BETTER_MEMORY_HOME; five tests previously called hook.main() with no home override and wrote to the developer's real memory.db/diagnostics/state files. - better_memory/hooks/contextual_inject.py: only open a sqlite connection when cfg.storage_backend == "sqlite" (nullcontext(None) otherwise, passed as memory_conn=None to build_backend, whose agentcore branch ignores it). Previously the hook unconditionally created memory.db (+WAL) on agentcore-only installs. _bump_diagnostic now also no-ops on a None conn. - Added test_agentcore_mode_does_not_open_sqlite_connection asserting connect() is never called and build_backend receives memory_conn=None when storage_backend=agentcore. - better_memory/services/relevant.py: _age_days now catches (ValueError, TypeError) from datetime.fromisoformat, not just ValueError.
![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
Reviewed all production code changes (config, migration 0012, contextual_inject hook, session_close, context_seen, reflection/relevant scoring, session_bootstrap, and the sqlite/agentcore/protocol storage backends). Signatures, new schema columns, and edge-case guards are all consistent, and no reachable logic errors, null-derefs, or resource issues were found in the added/modified lines.
No bugs were detected in this PR.
… rejected dict-unpack helper Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
There was a problem hiding this comment.
🟢 Claude BugBot Analysis
Reviewed the contextual-injection feature (hook, seen-store dedup, relevance scoring, session-bootstrap slimming, exposure-tracking protocol for sqlite/agentcore, migration 0012, and the rating directive). No genuine bugs found — all cross-file contracts (signatures, dataclass fields, SQL columns/CHECK constraints, config plumbing) verified consistent.
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
Redesigns how better-memory injects memories into running Claude Code sessions so the LLM actually attends to them, and closes the rating feedback loop for contextually-injected memories. Implements the approved spec
docs/superpowers/specs/2026-07-11-attention-first-injection-design.md(research-backed: recency-end injection, precision-over-recall floors, index+affordance over dumps, factual XML framing, Stop-hook rating directives).Bootstrap slimming (SessionStart)
BETTER_MEMORY_BOOTSTRAP_TOP_N, default 5;0= legacy full dump) rendered in full with full ids (fixes 8-char truncation bug that made semantic memories unrateable) and age stampsContextual injection (UserPromptSubmit / PreToolUse)
BETTER_MEMORY_CONTEXT_MIN_HITS, default 2) — below floor, inject nothingstate/context_seen_<session>.json) — a memory injects once per session (BETTER_MEMORY_CONTEXT_REINJECT_TURNSto loosen)<project-memory>XML block: factual framing, full ids, ages, dont-items flipped to corrective phrasing, inlinememory_creditaffordancesource='contextual'via newrecord_exposuresprotocol method (migration 0012; agentcore no-op per its no-exposure-log contract) — contextual injections are now rateable, closing the gap PR Contextual memory injection hook (UserPromptSubmit + PreToolUse) #76 leftRating loop
[bootstrap|retrieve|contextual]) and per-source countsDrive-by fix
os._exit(0)-killed the whole pytest run mid-suite with exit code 0 (looked like a phantom hang; explains the repo-root graveyard of truncated test logs). Allcreate_app()test call sites now passstart_watchdog=False.Testing
docs unaffected areas: observation-lifecycle.md, index.md, agentcore-setup.md (verified no drift)
🤖 Generated with Claude Code