hermes-learning-loop

Files (3)

• README.md
• SKILL.md
• learning-loop.js · scripts/

SKILL.md (first lines)

# Hermes Learning Loop for OpenClaw

**Inspired by**: [NousResearch/hermes-agent](https://github.com/NousResearch/hermes-agent) (Self-improving AI agent)

Implements Hermes Agent's core learning loop in OpenClaw — automatically extracts successful workflows, creates reusable skills, and persists curated knowledge across sessions.

## When to Use

- After completing complex multi-step tasks
- When you discover a workflow worth repeating
- Automatically via heartbeat (periodic nudge)
- When user says "remember this" or "save this workflow"

## Features

- **Periodic Nudge** — Auto-reflect every N tasks (default: 5)
- **Skill Extraction** — Convert successful workflows to reusable skills
- **4-Layer Memory** — Prompt memory + Session search + Skills + User modeling
- **Curated Memory** — Agent decides what's worth keeping (not logging everything)
- **Progressive Disclosure** — Only load skill summaries by default, full content on-demand
- **FTS5 Session Search** — SQLite-powered historical context retrieval

## Quick Start

### Install

```bash
clawhub install hermes-learning-loop
```

### Manual Trigger

```bash
# After completing a task
node learning-loop.js extract --session=<session_id>

# Create skill from workflow
node learning-loop.js create-skill --name="my-skill" --description="What it does"

# Periodic nudge (heartbeat)
node learning-loop.js nudge
```

### Auto-Trigger (Integration)

Add to HEARTBEAT.md:

```markdown
## Learning Loop - Periodic Nudge

**Frequency:** Every 5 tasks or 30 minutes

**Task:**
1. Run `node learning-loop.js nudge`
2. Review extracted memories
3. Approve/reject skill creations
```

## Architecture

### Learning Loop Cycle

```
┌─────────────────────────────────────────────────────────┐
│  1. Task Execution                                      │
│     - Agent completes task                              │
│     - Track tool calls, decisions, outcomes             │
└──────────────┬──────────────────────────────────────────┘
               │
               ▼
┌─────────────────────────────────────────────────────────┐
│  2. Periodic Nudge (every 5 tasks)                      │
│     - System prompt: "Reflect on recent activity"       │
│     - Agent evaluates: Is this worth persisting?        │
└──────────────┬──────────────────────────────────────────┘
               │
               ▼
┌─────────────────────────────────────────────────────────┐
│  3. Memory Extraction                                   │
│     - If valuable → Write to MEMORY.md / USER.md        │
│     - If workflow → Create skill file                   │
│     - If context → Index in session archive             │
└──────────────┬──────────────────────────────────────────┘
               │
               ▼
┌─────────────────────────────────────────────────────────┐
│  4. Skill Creation                                      │
│     - Extract steps, tool calls, file references        │
│     - Write to ~/.openclaw/skills/<category>/<name>/   │
│     - Follow agentskills.io specification               │
└─────────────────────────────────────────────────────────┘
```

### 4-Layer Memory System

| Layer | Purpose | Location | Load Timing |
|-------|---------|----------|-------------|
| **Prompt Memory** | Always-on context | `MEMORY.md`, `USER.md` | Every session start |
| **Skills** | Procedural memory (how-to) | `~/.openclaw/skills/` | On-demand (summary → full) |
| **Session Archive** | Episodic memory (what happened) | SQLite + FTS5 | Deliberate retrieval |
| **User Model** | Behavioral patterns | Optional (Honcho-like) | Continuous passive |

### Skill Triggers

Automatically create skill when:
- ✅ 5+ tool calls in sequence
- ✅ Recovered from error successfully
- ✅ User corrected → fixed approach
- ✅ Non-obvious workflow that worked
- ✅ Repeated pattern detected (3+ times)

## Usage Examples

### Example 1: Post-Task Skill Extraction

```bash
# After completing complex deployment
export SESSION_ID="2026-04-03-deploy"
export TASK_OUTCOME="success"

From Arry8/openclaw-edge, MIT.

More from Arry8/openclaw-edge