Hermes Agent Curator Guide: FIX Your Agent Skills!

Hermes Curator auto-prunes your agent's skill catalog, but you'll likely hit a config gap and a silent daemon crash on first run.

What Curator actually does

Curator is a background maintenance layer for agent-created skills in Hermes Agent. The self-improvement loop lets your agent create new skills mid-conversation, and over time Banner's agent had accumulated 43 agent-created skills — with the top five least recently used showing use never. Curator consolidates near-duplicate skills via clustering, archives unused ones, tracks last updated and last created, and can be scheduled. On Banner's first run it archived 27 skills.

The two commands that matter

• hermes curator status — confirms it's enabled and shows your skill count.
• hermes curator run sync — streams the log; hermes curator run runs in the background.

A successful run writes a timestamped directory under .hermes/logs/curator/ containing run.json (machine-readable) and report.md (human-readable). The LLM final summary block is what to read — it lists clusters, patched/absorbed near-duplicates, skill demotions, and deliberately-skipped clusters (bundled, native skills are never touched).

Recovery and the archival gotcha

Curator never auto-deletes — worst case is archival. Banner lost skills he wanted archived and had to run hermes curator restore <name>. If a skill you need is archived, that command brings it back.

The two bugs Ron hit on Loki

1. Missing config block. Loki's ~/.hermes config had no Curator block, so the curator folder never got created in .hermes/logs/. Feed your agent the official Curator docs and ask it to verify config before anything else — don't let it just mkdir the folder, or it will hallucinate and make up information. 2. Silent daemon crash. The LLM review phase runs in a daemon thread; the previous five runs hit no module named fire when spawning the forked AI agent, the thread died silently, and the LLM pass was never completed. The first run took about an hour and produced no report — a 15–20 minute run is closer to a healthy baseline.

Verdict for builders

The feature is worth turning on, but budget an hour of debugging for the config + daemon thread issue. Bundled skills are safe; only agent-created skills are at risk, and only of being archived.