Memory

/memory — Memory Corpus Quality

Read-only rollup of how the memory corpus is doing: are candidates being captured, are they being approved, is the corpus being read, are the hooks healthy. Distinct from /reflect (which applies/dismisses individual candidates) and /recall (which queries memory) — this is the audit view, like /infra-health for hooks and /stats for cost.

Commands

/memory health — corpus quality rollup

python3 ~/.claude/skills/memory/scripts/memory_health.py           # text report
python3 ~/.claude/skills/memory/scripts/memory_health.py --json    # one JSON object

Prints a read-only report across seven sections + a one-line verdict. Never mutates the corpus, the review queue, or the logs; degrades gracefully on every missing source (prints "(no data yet)", never errors).

Sections:

• Capture — memory-infer hook: run count, total candidates yielded (sum of appended=N from hook detail), yield distribution (0/1/2/3+ per run).
• Approval rate — over resolved queue entries: applied / (applied+dismissed), pending tracked separately. "(no resolves yet)" until the first candidate is applied or dismissed.
• Invalidations — queue kind=invalidation count, split stale-90d vs overlap (parsed from description).
• Utilization — from memory-consulted.json: % of corpus never consulted, top-5 consulted, count consulted in the last 30d. "(no data yet — read hook not yet active)" when the sink is absent.
• Corpus shape — total / no-frontmatter / flat-schema / canonical; type distribution; confidence distribution; last_verified age histogram (none / <30d / 30-90d / >90d); superseded_by count.
• Hook health — per memory hook: run count, error rate, mean duration.
• Issues — never-consulted>60d, fastest-staling (high-confidence by age, top 5), zero-yield infer sessions, schema-adoption %.

Verdict — memory: HEALTHY or memory: NEEDS ATTENTION (N flags). Flags fire when: any hook error rate >10%, canonical adoption <10%, or never-consulted >50%.

/memory review — propose harness tweaks

python3 ~/.claude/skills/memory/scripts/memory_review.py

Reads the /memory health findings and proposes concrete, accept/defer harness tweaks — each as finding → why → action. Advisory and deterministic: it changes nothing. Accepting a tweak means making it a follow-up task.

Typical proposals: low schema adoption → run the S3 backfill; high never-consulted % → archive cruft / reseed warm-start; >25% zero-yield infer sessions → raise the turn threshold or tighten the prompt; low approval rate → tighten the prompt; hook error rate >10% → inspect the hook log; staling memories → re-verify via /reflect.

Data sources (all read-only)

--json emits one object {capture, approval, invalidations, utilization, corpus, hooks, issues, verdict} for the downstream /memory review command.