self-improving-agent

Self-Improving Agent — Continuous Learning and Improvement

Overview

This skill enables OpenClaw agents to continuously learn from experience by:

1. Logging errors and corrections — Building a knowledge base of what works and what doesn't
2. Capturing learnings — Documenting successful patterns and solutions
3. Updating preferences — Remembering user preferences and context
4. Autonomous improvement — Using Evolver meta-skill for self-optimization

How It Works

Automatic Activation

The skill activates automatically when:

• An error occurs during task execution
• A successful solution is found after multiple attempts
• User provides correction or feedback
• Agent discovers a non-obvious solution

Learning Capture Format

timestamp: 2026-04-01T09:14:00Z
context: "Web application development task"
situation: "Tried to install package with wrong Python version"
error: "Package requires Python 3.10+, have 3.9"
solution: "Used pyenv to switch to Python 3.11, then installed"
learning: "Always check Python version requirements before installation"
category: "environment_setup"
priority: "medium"

Memory Integration

Learnings are stored in:

• ~/.openclaw/workspace/memory/learnings/ — Categorized learning files
• ~/.openclaw/workspace/memory/preferences.md — User and task preferences
• Session context — Immediate recall during current session

Integration with Evolver Meta-Skill

This skill works with the Evolver meta-skill to create a complete self-improvement system:

Raw Experience
    ↓
Self-Improving Agent (Capture)
    ├── Log errors and corrections
    ├── Document successful patterns
    └── Update user preferences
    ↓
Evolver Meta-Skill (Analyze & Improve)
    ├── Analyze patterns across learnings
    ├── Identify optimization opportunities
    └── Implement systematic improvements
    ↓
Enhanced Agent Performance

Setup for Different Agents

OpenClaw

• Activation: Workspace injection + inter-agent messaging
• Setup: Automatic via skill loading
• Detection: Via session tools and workspace files

Claude Code

• Activation: Hooks (UserPromptSubmit, PostToolUse)
• Setup: .claude/settings.json with hook configuration
• Detection: Automatic via hook scripts

Codex

• Activation: Hooks (same pattern as Claude Code)
• Setup: .codex/settings.json with hook configuration
• Detection: Automatic via hook scripts

GitHub Copilot

• Activation: Manual (no hook support)
• Setup: Add to .github/copilot-instructions.md
• Detection: Manual triggering

Usage Examples

Example 1: Error Learning

Situation: Failed to parse JSON response from API
Action: Log error pattern and solution
Result: Future API calls include better error handling

Example 2: Preference Learning

Situation: User prefers markdown tables over bullet lists
Action: Update user preferences
Result: Future responses use markdown tables for similar data

Example 3: Pattern Recognition

Situation: Successfully solved similar problem 3 times
Action: Document pattern as reusable skill
Result: New micro-skill created for future use

File Structure

~/.openclaw/workspace/
├── memory/
│   ├── learnings/
│   │   ├── api_errors.md
│   │   ├── environment_setup.md
│   │   ├── user_preferences.md
│   │   └── successful_patterns.md
│   └── preferences.md
└── skills/
    └── self-improving-agent/
        └── SKILL.md

Configuration

# Optional configuration in workspace
self_improvement:
  # What to capture
  capture_errors: true
  capture_successes: true
  capture_preferences: true
  
  # Storage
  learning_retention_days: 90
  max_learnings_per_file: 100
  
  # Privacy
  exclude_sensitive: true
  hash_identifiable_info: true
  
  # Integration
  enable_evolver_integration: true
  improvement_check_frequency: "weekly"

Quick Start

Manual Triggering

# Log a learning manually
echo "After completing this task, evaluate if any learnings should be logged using the self-improvement skill format." | openclaw

# View all learnings
cat ~/.openclaw/workspace/memory/learnings/*.md

# Check improvement status
openclaw skills list | grep -i improve

Automatic Operation

Once installed, the skill automatically:

1. Monitors task execution for learnable moments
2. Captures errors, corrections, and successes
3. Updates memory with new learnings
4. Integrates with Evolver for systematic improvement

Best Practices

1. Be specific — Include context, error, solution, and learning
2. Categorize — Use consistent categories for easy retrieval
3. Prioritize — Mark high-value learnings for immediate attention
4. Review regularly — Periodically review and consolidate learnings
5. Share learnings — In multi-agent teams, share relevant learnings

Safety and Privacy

• No sensitive data — Automatically filters passwords, keys, PII
• Local storage only — All learnings stay on local machine
• User control — Users can review and delete learnings
• Opt-out available — Can be disabled per session or globally

---

"The only true wisdom is in knowing you know nothing — and remembering what you learn along the way."