Curator

Harvest writes; curator keeps it trustworthy. An unmaintained knowledge base rots — contradictions accumulate, duplicates fragment retrieval, and references outlive the things they point to. Two passes, run together or alone.

Lint pass

Routine-tier work: mechanical checks, conservative fixes.

1. Scope the window. Find the last curator-run memory (nmem memories search "curator run"); lint what changed since. First run: sweep everything (nmem memories list, nmem stats).
2. Duplicates — search topic clusters the window touched; near-identical entries merge into the strongest one (update), losers get supersede links.
3. Contradictions — where a newer memory disagrees with an older one, verify which is true now; deprecate the loser with the reason recorded, never silently.
4. Dead references — memories naming files, flags, skills, or commands: verify each still exists (check the repo or filesystem). Update what moved; deprecate what vanished.
5. Expired — entries whose stated expiry has passed: re-verify or deprecate. Time-sensitive claims with no expiry get one added now.
6. Harvest gaps — list the window's threads (nmem threads list) and compare against memories added in the same span; a non-trivial session that saved nothing is a harvest miss. Read the thread, save what a future session would need, or confirm there was genuinely nothing to keep. Threads import is manual — a sparse window may mean unsynced, not idle; say so in the report instead of guessing.

Fix verbs, in order of preference: update (correct in place) → supersede (old links to replacement) → deprecate (obsolete, kept for audit) → delete (only for entries wrong from birth). History stays legible.

Synthesis pass

Premium-tier work: judgment across the whole window.

Read the window's memories plus the feed (nmem feed), then write a short report to the user: what entered, what themes are forming, what drifts from earlier decisions, what deserves attention next. This is the pass that surfaces cross-source connections no single entry shows.

Closing a run

End every run by saving one memory: curator run <date> — <n> linted, <fixes>, <one-line synthesis>. That entry is the next run's starting line.

Run manually until the output proves itself; then schedule it (weekly is the right default) via the runtime's scheduler.