Hermes Agent — Learning-Loop Skill
> Turn OpenClaw into a learning-loop agent with seeded workspace rules, skill promotion, reflective memory, and proactive maintenance.
When to Use
User wants OpenClaw to feel more persistent, more self-correcting, and more proactive across sessions. Agent handles workspace seeding, reflective memory, repeated-workflow capture, and promotion of reusable patterns into future skills or stable rules.
Architecture
Memory lives in ~/hermes-agent/. If ~/hermes-agent/ does not exist, run setup.md. See memory-template.md for structure.
~/hermes-agent/
|-- memory.md # HOT: current rules, active signals, stable lessons
|-- promotions.md # Candidate workflows that may graduate into skills
|-- reflections.md # Recent post-task reflections
|-- workspace-state.md # Which OpenClaw files were extended and how
`-- archive/ # Cold lessons and retired patterns
Quick Reference
| Topic | File |
|---|---|
| Setup guide | setup.md |
| Memory template | memory-template.md |
| Loop design | loop.md |
| OpenClaw seed blocks | openclaw-seed.md |
| Skill promotion rules | promotion.md |
Core Rules
1. Seed OpenClaw Non-Destructively
- Extend
AGENTS.md,SOUL.md, orHEARTBEAT.mdonly with additive blocks. - Never replace the whole file, remove unrelated lines, or rewrite the user's persona.
2. Retrieve Before Non-Trivial Work
- Before any multi-step, failure-prone, or repeated workflow, read
~/hermes-agent/memory.md.
3. Reflect Immediately After Significant Work
4. Promote Repetition Into Stable Capability
5. Keep Memory Bounded and Operational
- Prefer
AGENTS.mdfor routing rules,SOUL.mdfor tone pressure, andHEARTBEAT.mdfor periodic maintenance. Do not duplicate the same rule in all three.
6. Respect Local Boundaries
- Never store credentials, secrets, payment data, health data, or copied transcripts.
- Never modify
SKILL.mdor claim native OpenClaw hooks that do not exist.
Common Traps
- Treating Hermes as branding only → OpenClaw sounds smarter, but behavior does not compound.
- Rewriting whole workspace files → destroys user custom context and creates trust debt.
- Logging every tiny thought → memory becomes noisy and retrieval quality drops.
- Promoting a one-off fix to a global rule → future sessions inherit the wrong behavior.
- Creating a new skill too early → the user gets premature complexity instead of a refined workflow.
Security & Privacy
Data that stays local:
- Additive seed blocks placed in local OpenClaw workspace files
Data that leaves your machine:
- None by this skill itself
This skill does NOT:
- access files outside the local OpenClaw workspace and ~/hermes-agent/
- replace the full contents of AGENTS.md, SOUL.md, or HEARTBEAT.md
