Meta-Memory Review

Purpose: periodic self-optimization of the memory STRUCTURE (not its content). A strong-model pass reads recent session trajectories + recurring-failure signals and PROPOSES concrete edits to the memory rules / schema / tiers. Proposals are surfaced to the operator; NEVER auto-applied (memory-rule/schema/tier edits are CONFIRM-gated).

Lift provenance: AutoMem (arXiv 2607.01224) loop-1 = "a strong LLM reviews complete agent trajectories and iteratively revises the memory structure." Take loop-1 (structure-revision-by-review) and SKIP loop-2 (RL fine-tuning, impossible on a frozen model). The problem it targets: "a single memory mistake hides long before it surfaces" = recurring drift / over-hardening / noop-loop pains, which more text-lessons have failed to fix.

Inputs (READ-ONLY)

1. logs/session.jsonl, last ~7 days of session events (categories, save extractions, failures, event mix).
2. memory/Learnings.md, tail + scan for RECURRENCE markers ("Nx", "N-th occurrence", "3rd/4th instance"). Repeated failure modes are the highest-signal targets.
3. logs/errors.jsonl + skill-failure tracker (scripts/eval/skill_failure_tracker.py output, or its store), skills failing 3x/14d.
4. reality-feedback ledger (reality_feedback.sqlite if present), verdicts + friction signals.
5. The CURRENT rules to propose AGAINST: skills/save/SKILL.md (extraction rules), memory/MEMORY.md (tier-1 index + bloat), frontmatter conventions across memory/Context|Infra|Feedback.

Review lenses (produce proposals under each; SKIP a lens with no evidence)

• A. Extraction gaps: patterns that recurred in sessions but save.md did NOT capture, or mis-categorized. → propose a save.md rule edit.
• B. Schema: frontmatter/format inconsistencies or a missing field that would improve recall/dedup. → propose the schema change.
• C. Tier misassignment: tier-1 (always-loaded) entries that are stale / low-value (every line is reply-cost tax), or tier-2 items that should be promoted. → propose move/prune (cross-check MEMORY.md detox + reaper history).
• D. Recurring-failure → STRUCTURAL guard (the AutoMem core): for each Learnings recurrence (e.g. over-hardening 4x, continue-turn-drop 3x), ask: is there a memory/skill STRUCTURE change (hook, gate, schema field, a load-order fix) that would actually guard it, versus appending an (N+1)th text lesson that already failed N times? Propose the structural fix.

Output + GATE

• If substantive proposals found: write the full review to outputs/raw/agent/YYYY-MM-DD-meta-memory-review.md (frontmatter status: raw), then surface a CONCISE summary to the operator: the top 3-5 proposals, each = observed pattern + the specific edit + why. Ask which to apply.
• NEVER auto-apply. Only after the operator approves specific proposals are they edited in.
• SILENT ON CLEAN: if no substantive proposal (a quiet week), log ONE line to session.jsonl and do NOT ping the operator. Evidence-gated (per the "no evidence-on-clean-path" rule). No noise.

Cadence

Weekly cron meta-memory-review-weekly (Sunday 05:00 local, strong model / high effort). Also manual via /meta-memory-review.