agentsclimarketplace
Sign in

Memory lint

Skill bentleypark/aiwatch/.claude/skills/memory-lint

Periodic health check of the file-based LLM Wiki memory (the `~/.claude/.../memory/` bundle) — the "lint" operation of the Karpathy LLM-Wiki pattern. Invoke every week or two, or after a batch of memory-ingest writes, or when the memory "feels stale". Finds contradictions, stale/superseded claims, orphan pages, broken `[[wikilinks]]`, index drift (file↔MEMORY.md mismatch), unquoted-`#` frontmatter truncation, and redundant clusters worth merging — then records findings in log.md. Read-only proposals for anything destructive; the user confirms merges/deletions.From its SKILL.md

Install
npx -y skills add bentleypark/aiwatch --skill memory-lint

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

SKILL.md

4.1 KB, as published. Nobody here has run it

memory-lint — keep the LLM Wiki healthy

The file memory (~/.claude/projects/-Users-bentley-Desktop-bentely-aiwatch-aiwatch/memory/) compounds over time; without periodic linting it drifts. This is the lint operation of the LLM-Wiki pattern (#891).

When to run

  • Every 1–2 weeks, after a batch of memory-ingest writes, or when recall feels noisy/stale.

Checks

  1. Index drift — every memory/*.md (except MEMORY.md) has exactly one MEMORY.md line, and every MEMORY.md link resolves to a file. Report NOT INDEXED / MISSING FILE.

  2. Broken wikilinks — every [[stem]] resolves to an existing memory/stem.md (wikilinks target the type_-prefixed filename stem, not the frontmatter name; or a deliberate forward-marker). Report danglers.

  3. Frontmatter integrity — every page has name + type; no description/title truncated by an unquoted bare # (grep for descriptions containing # without surrounding quotes).

  4. Contradictions / stale claims — two pages that assert opposite things, or a page whose claim a later session invalidated (e.g. a file/flag it names no longer exists — verify against current code before flagging).

  5. Redundant clusters — near-duplicate pages on one topic (e.g. the #891 Phase 1 review/close/verify clusters) that should be merged into one, preserving each original's Why.

  6. Decision-graph structure (#967) — run npm run lint:graph -- --github. The -- is required: npm run lint:graph --github silently swallows the flag and runs the offline subset, which then passes while skipping every liveness check. (node scripts/lint-decision-graph.mjs --github works too.) It checks what is mechanical and refuses to guess at what is not:

    • edge grammar — bounds:: only on a constraint_* page and only ever targeting a decision page; constrains:: only on decision_*; advances:: only on initiative_* (rule 1b); blocks:: never a wiki edge.
    • dead edges (--github) — every advances:: #N resolves to an OPEN issue that is not a PR.
    • duplicate claims — no issue is a slice of two initiatives.
    • dangling [[wikilinks]] — over the WHOLE bundle, not just the graph pages: this is the only mechanical version of check 2 (which is otherwise a prose instruction), so a broken link anywhere is a hard finding here.
    • unclaimed candidates (--github) — open area:biz/area:marketing issues that no initiative claims. A report, never a finding, and exit code 0. Whether an issue advances an initiative is a judgement; a lint that decides it manufactures false confidence. Read the list, decide, edit the page.

    Structural findings exit non-zero. The lint cannot run in CI — the memory bundle is harness-global, so Actions has nothing to check out — but its pure functions are CI-gated via npm run test:scripts.

Output

  • Write a YYYY-MM-DD lint: <n> findings (<summary>) line to memory/log.md.
  • Auto-fix only the safe, mechanical issues (repoint a broken wikilink to its renamed target, add a missing index line, quote a #-truncated description).
  • Propose — do NOT auto-apply — anything lossy (merging pages, deleting stale claims). Present the plan and wait for user confirmation (user feedback is never deleted unilaterally; merge preserves the Why).

Relationship

Pairs with memory-ingest (the write side). Neither touches the repo — the memory dir is harness-global.

Keep looking

Skills are one crate of 627,801. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.