Hermes Learning Loop: Memory & Skills

Why this page exists: Hermes’s differentiator is not “yet another Telegram bot.” It is that the agent is designed to get more capable the longer it runs. If you leave memory/skill creation off, you are mostly buying a gateway + tools stack—not the product thesis.

Loop components (operator map)

Component	What you should notice	Risk if ignored
Persistent memory	Preferences, projects, environment facts survive sessions	Stale or over-retained sensitive data
Autonomous skill creation	After hard tasks, SKILL.md-style procedures appear	Unreviewed “code-like” instructions accumulate
Skill self-improvement	Skills evolve when reused	Silent behavior drift
FTS5 + summarization	Cross-session recall that is searchable	Extra token spend on consolidation jobs
User modeling (optional)	Deeper personalization (e.g. Honcho)	Higher sensitivity of stored profile

Skills: install vs invent

Hermes participates in the agentskills.io open skill format and documents a Skills Hub / browse-install flow. Typical operator commands (confirm names in your build):

hermes skills browse
hermes skills search k8s
hermes skills install <skill-ref>   # docs often mention a security scan before install

Two sources of skills:

Installed / hub skills — treat like packages; prefer sources you trust.
Agent-authored skills — treat like untrusted code that happens to be Markdown + instructions. Review before production.

Mindset borrowed from our OpenClaw guide: skills audit checklist (adapt the questions to Hermes paths under ~/.hermes/skills/).

Memory hygiene

Know where memory/session files live; include them in backup and in incident wipe procedures.
Do not paste production secrets into chat “so the agent remembers.” Use env/config.
Schedule periodic review: delete wrong memories; archive skills you no longer want.
If you enable dialectic/user-modeling backends, treat the profile as PII.

Cost reality of self-improvement

Memory consolidation, skill generation, and summarization are background token consumers. Budget them like cron jobs:

Use cheaper models for consolidation if the product allows routing.
Cap how often nudges/cron fire during the first month.
Cross-check spend patterns with our cost playbook (written for OpenClaw but the hosting vs tokens split still applies).

SOUL.md & context files

Hermes docs highlight personality via global SOUL.md and project context files that shape conversations. Start minimal: one clear personality + one project context per repo. Overlong soul files burn tokens and fight the skill system.

When to turn the loop off

Shared team bots where self-authored skills would leak process knowledge.
Compliance environments that forbid unsupervised “agent-written procedures.”
Debugging weird behavior—freeze skill writes, reproduce, then re-enable.

Blank Slate / disabled toolsets (see install guide) help you prove baseline behavior without the loop.

Hermes column

Last updated: 2026-07-14 · Independent operator notes on OpenClaw Roadmap. Verify commands and features on the official Hermes docs and GitHub—this space moves fast.