Hermes Curator: An Autonomous Skill-Maintenance Agent

Build Log — v0.12.0 (2026.4.30): The “Curator” release introduced an autonomous background agent that maintains your skill library. 1,096 commits, 550 merged PRs, and 213 community contributors went into this release [1]. This post walks through what the Curator is, how to configure it, and what happens during a curation cycle.

Why a Curator?

Skills are procedural memory — markdown documents in ~/.hermes/skills/ that the agent reads as part of its system prompt. Over time, your skill library accumulates:

• Drafts — one-off workflows that should probably be merged
• Stale skills — procedures for tools or APIs you no longer use
• Duplicates — two skills covering the same task with slightly different approaches
• Dead references — scripts or templates that were deleted but still linked in a skill’s frontmatter

Manually cleaning this up is tedious. The Curator automates it.

Architecture

The Curator runs as a background agent on the gateway’s cron ticker:

Gateway cron ticker (7-day default)
  └─ auxiliary.curator process
       ├─ Scans ~/.hermes/skills/ (excluding bundled/hub skills)
       ├─ Grades each skill: usage stats, content quality, freshness
       ├─ Classifies: keep / consolidate / prune
       ├─ If consolidate → merges related skills, updates references
       ├─ If prune → archives dead skills (never deletes)
       └─ Writes report to logs/curator/run.json + REPORT.md

It inherits the parent gateway’s config — same provider, model, and credentials. Pinned skills and hub-installed skills are protected.

Setting It Up

The Curator is unified under auxiliary.curator. Enable it in your config:

# ~/.config/hermes/config.yaml
auxiliary:
  curator:
    enabled: true
    cycle_days: 7  # default; change to 14 for less frequent runs
    model: auto    # inherit from primary, or pin a specific model

Or set it via CLI:

hermes config set auxiliary.curator.enabled true
hermes config set auxiliary.curator.cycle_days 14

Pick which model it uses:

hermes model          # select main model
# or pin a specific model just for the curator:
hermes config set auxiliary.curator.model gpt-4o

Checking Status

See how the curator ranks your skills:

hermes curator status

Output (example):

Most-used skills (last 30 days):
  ↔ github-pr-workflow         — 24 uses
  ↔ static-blog                — 19 uses
  ↔ content-pipeline           — 15 uses

Least-used skills (last 30 days):
  ↔ legacy-deploy-script       — 0 uses
  ↔ old-mcp-setup              — 0 uses

What a Curation Run Looks Like

When the Curator fires, it writes two outputs to logs/curator/:

run.json — machine-readable summary:

{
  "run_id": "cur-20260526-001",
  "started_at": "2026-05-26T03:00:00Z",
  "skills_scanned": 47,
  "skills_pinned": 5,
  "consolidated": 3,
  "pruned": 2,
  "duration_seconds": 142
}

REPORT.md — human-readable decisions:

# Curator Run — 2026-05-26

## Consolidated
- `deploy-guide-v1` + `deploy-guide-v2` → `deploy-guide`
- `troubleshooting-docker` + `docker-common-issues` → `docker-troubleshooting`

## Pruned (archived → ~/.hermes/trash/)
- `obsolete-model-config` — provider no longer supported
- `v0.8-migration` — migration was one-time, 3 months stale

## Protected
- 5 pinned skills (not evaluated)
- 12 hub-installed skills (not evaluated)

Consolidation Logic

The consolidation step is surprisingly sophisticated:

1. Content overlap scoring — the agent compares skill bodies for structural similarity (steps, commands, pitfalls)
2. Usage check — two low-usage skills covering the same domain get merged; high-usage skills with overlapping content get suggestions instead
3. Reference reconciliation — merged skills get unified reference paths; old paths are forwarded via absorbed_into metadata
4. Version bump — the merged skill gets version: 2.0.0

Guardrails

The Curator has safety boundaries:

Guardrail	What it prevents
Pinned skills	skill_manage(action='delete') refused on pinned skills (protected)
Hub skills	Skills installed via hermes skills install <hub-url> are never modified
absorbed_into tracking	When consolidating, sets metadata so downstream consumers update references
Trash directory	Pruned skills go to ~/.hermes/trash/ not /dev/null — recoverable with mv
Run report always written	Even if no actions taken, report documents the scan for audit trail

What This Means for Your Workflow

Before the Curator, skills were write-once, maybe-update-never. After a few months, you’d have 80+ skill files, half of which were drafts, dead ends, or superseded by better approaches.

Now the library stays lean:

• Skills that prove their value get used more → higher ranking → survive curation
• Merged skills collapse surface area → faster skill_view loads
• Archived junk is out of sight → skills_list output stays scannable
• Each run produces a report → you know exactly what changed

Gotchas

A few things I hit setting this up:

Cycle timing. The 7-day cycle starts from gateway boot, not from when you enable it. If you hermes config set at noon, the first run happens at noon 7 days later (assuming the gateway stays up). To force an immediate run, restart the gateway.

Model choice matters. If you pin a weak model for the curator to save tokens, you’ll get worse consolidation decisions. gpt-4o or equivalent is recommended. The curator’s runtime is cheap — in practice, a single scan costs about the same as 2-3 normal prompts.

Standalone mode. You can also run the curator ad-hoc with hermes curator run (v0.12.1+). This bypasses the cron cycle and runs immediately in the foreground with streaming output.

Summary

The autonomous Curator takes skill maintenance from a manual chore to a hands-off background service. Combined with the upgraded self-improvement loop (rubric-based review forks that handle references/ and templates/ sub-files), Hermes Agent now maintains its own knowledge base without intervention.

If you’re running Hermes Agent v0.12.0 or later, the Curator is already enabled by default — just run hermes curator status to see what it’s been up to.

References

[1] Hermes Agent v0.12.0 Release — GitHub [2] Official release notes for v0.12.0 — RELEASE_v0.12.0.md [3] Hermes Agent Changelog — hermes-ai.net [4] Skills documentation — Hermes Agent docs [5] Hermes Agent GitHub — NousResearch