Hermes AgentHermes智能体

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.

CODEBLOCK0

Quick Reference

Core Rules

1. Seed OpenClaw Non-Destructively

• - Extend AGENTS.md, SOUL.md, or HEARTBEAT.md only with additive blocks.
• Never replace the whole file, remove unrelated lines, or rewrite the user's persona.
• If a seed block already exists, refine only the smallest relevant section.

2. Retrieve Before Non-Trivial Work

• - Before any multi-step, failure-prone, or repeated workflow, read ~/hermes-agent/memory.md.
• Then read at most one extra Hermes support file unless the task clearly needs more.
• Do not load the full Hermes stack "just in case".
• Skip Hermes retrieval for trivial one-shot replies, small factual answers, or casual chat.

3. Reflect Immediately After Significant Work

• - After meaningful execution, compare intent, outcome, and friction.
• Write one concise reflection to reflections.md when the lesson is reusable.
• If the lesson changes future behavior, also distill it into memory.md.

4. Promote Repetition Into Stable Capability

• - If the same workaround, pattern, or procedure succeeds three times, log it in promotions.md.
• If the pattern is broad and reusable, recommend turning it into a dedicated skill.
• If the pattern only sharpens current behavior, keep it as a workspace rule instead.

5. Keep Memory Bounded and Operational

• - memory.md stays short, current, and execution-oriented.
• Move stale or superseded lessons to archive/.
• Prefer one strong rule over five similar notes.
• Prefer AGENTS.md for routing rules, SOUL.md for tone pressure, and HEARTBEAT.md for periodic maintenance. Do not duplicate the same rule in all three.

6. Respect Local Boundaries

• - Store only operational lessons, preferences, and workflow decisions.
• Never store credentials, secrets, payment data, health data, or copied transcripts.
• Never modify SKILL.md or 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:

• - Lessons, reflections, workspace integration state, and promotion candidates under INLINECODE12
• Additive seed blocks placed in local OpenClaw workspace files

Data that leaves your machine:

• - None by this skill itself

This skill does NOT:

• - make network requests
• access files outside the local OpenClaw workspace and INLINECODE13
• replace the full contents of AGENTS.md, SOUL.md, or HEARTBEAT.md
• store secrets or sensitive personal data

Related Skills

• - self-improving - capture corrections and recurring lessons so execution quality compounds
• INLINECODE16 - structure durable local memory for agent continuity
• INLINECODE17 - formalize repeated operating patterns into stable execution sequences
• INLINECODE18 - turn a proven repeated workflow into a dedicated skill package

Feedback

• - If useful: INLINECODE19
• Stay updated: INLINECODE20