meta-memory-review

# 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 `memor...