Self-Improving Agent

A continuous learning agent system for OpenClaw that learns from:

• ✅ Interactions - Learn from every conversation
• ✅ Errors - Learn from mistakes to avoid them
• ✅ Recoveries - Learn what works to recover successfully

✨ Features

• 🧠 Continuous Learning - Learns from every interaction
• ❌ Error Learning - Learns from mistakes to avoid repetition
• ✅ Recovery Learning - Learns successful recovery patterns
• 🔄 Auto-Improvement - Automatically applies improvements
• 📚 Memory System - Stores and retrieves learnings
• 🔌 Hook System - Extensible hook system for custom improvements
• 📊 Progress Tracking - Track improvement over time
• 🔧 Python CLI - Easy to use command-line interface
• 🧩 OpenClaw Skill - Seamless integration with OpenClaw

🚀 Quick Start

Prerequisites

• Python 3.10+
• OpenClaw installed
• pip or uv package manager

Installation

As OpenClaw Skill

# Clone to OpenClaw skills directory
cd ~/.openclaw/workspace/skills
git clone https://github.com/leohuang8688/self-improving-agent.git

Auto-learning is enabled by default! After each OpenClaw session, the agent will automatically learn and improve.

As Python Package

# Clone the repository
git clone https://github.com/leohuang8688/self-improving-agent.git
cd self-improving-agent

# Install with pip
pip install -e .

# Or install with uv
uv pip install -e .

Basic Usage

Automatic Mode (Recommended)

Just use OpenClaw normally! Learning happens automatically after each session, error, or recovery.

# Start OpenClaw
openclaw

# Use it normally...

# Exit OpenClaw
# → Auto-learning triggers automatically!

Manual Mode

# Run the self-improving agent
python -m self_improving_agent run

# Learn from last session
python -m self_improving_agent learn

# Learn from errors
python -m self_improving_agent learn-errors

# Learn from recoveries
python -m self_improving_agent learn-recoveries

# Review all learnings
python -m self_improving_agent review

# Export learnings to file
python -m self_improving_agent export

📖 Commands

run - Run the Agent

Executes the self-improving agent with all applied improvements.

python -m self_improving_agent run --workspace /path/to/workspace

learn - Learn from Session

Analyzes the last session and extracts learnings.

python -m self_improving_agent learn --verbose

learn-errors - Learn from Errors

Analyzes error logs and extracts learnings.

python -m self_improving_agent learn-errors --verbose

learn-recoveries - Learn from Recoveries

Analyzes recovery logs and extracts successful patterns.

python -m self_improving_agent learn-recoveries --verbose

review - Review Learnings

Reviews all stored learnings.

python -m self_improving_agent review --verbose

export - Export Learnings

Exports all learnings to a markdown file.

python -m self_improving_agent export

🪝 Hook System

Available Hooks

onSessionEnd(session)

• Triggered: When session ends
• Purpose: Post-session learning, cleanup, save state
• Parameter: session - Session object

onError(error)

• Triggered: When an error occurs
• Purpose: Error logging, learning from mistakes
• Parameter: error - Error object

onRecovery()

• Triggered: When recovering from error
• Purpose: Learn successful recovery patterns
• Parameter: None

📝 Learning Types

1. Session Learning

• Triggered by: onSessionEnd
• Learns from: Conversation patterns, user preferences
• Stored in: learnings/sessions.json

2. Error Learning

• Triggered by: onError
• Learns from: Mistakes, errors, failures
• Stored in: learnings/errors.json
• Purpose: Avoid repeating mistakes

3. Recovery Learning

• Triggered by: onRecovery
• Learns from: Successful recoveries
• Stored in: learnings/recoveries.json
• Purpose: Repeat successful recovery patterns

📁 Project Structure

self-improving-agent/
├── src/
│   ├── hooks.py              # Hook manager
│   └── ...
├── hooks/
│   └── error_learning.py     # Error learning hook
├── learnings/
│   ├── sessions.json         # Session learnings
│   ├── errors.json           # Error learnings
│   └── recoveries.json       # Recovery learnings
├── main.py
├── README.md
└── SKILL.md

🔧 Configuration

In OpenClaw configuration file:

{
  "hooks": {
    "internal": {
      "enabled": true,
      "entries": {
        "self-improve": {
          "enabled": true
        },
        "error-learning": {
          "enabled": true
        }
      }
    }
  }
}

📊 Learning Files

learnings/errors.json

[
  {
    "timestamp": "2026-03-15T12:30:00",
    "error_type": "ValueError",
    "error_message": "Invalid parameter",
    "context": {
      "traceback": "..."
    }
  }
]

learnings/recoveries.json

[
  {
    "timestamp": "2026-03-15T12:35:00",
    "recovery_method": "automatic_recovery"
  }
]

🎯 Best Practices

1. Review learnings regularly - Review and apply learnings weekly
2. Export learnings - Export learnings for backup
3. Apply learnings - Apply learnings to improve future performance
4. Clean old learnings - Remove outdated learnings periodically

📝 License

MIT License

👨‍💻 Author

PocketAI for Leo - OpenClaw Community