Hermes Self-Learning Agent

What It Solves

A lot of OpenClaw agents work fine for a single conversation, then behave like a new hire in the next one. They do not remember which repeated workflows failed before, nor do they turn repeated fixes into stable rules. Hermes Self-Learning Agent addresses this lack of compounding behavior by writing experience back into local workspace files, so later sessions can retrieve, reuse, and correct it instead of guessing again.

How It Works

The skill treats ~/hermes-agent/ as its local memory directory and uses setup.md, memory-template.md, loop.md, openclaw-seed.md, and promotion.md to form a learning loop:

• Non-destructive seeding: it appends blocks to AGENTS.md, SOUL.md, or HEARTBEAT.md instead of rewriting whole user files.
• Retrieve before non-trivial work: for multi-step, failure-prone, or repeated workflows, it reads ~/hermes-agent/memory.md first.
• Reflect after significant work: it records useful conclusions and failure points immediately after the task.
• Promote gradually: repeated patterns become stable rules or candidate skills before they are promoted too early.
• Separate responsibilities: routing rules stay in AGENTS.md, tone constraints in SOUL.md, and periodic maintenance in HEARTBEAT.md, avoiding duplicate rules across all three.

The boundaries are explicit: it does not store credentials, secrets, payment data, health data, or copied transcripts, and it does not claim non-existent native OpenClaw hooks. It behaves more like a local learning protocol than a cloud memory service or a branding layer.