Self-Improving Agent

Overview

This skill enables OpenClaw to continuously improve by capturing learnings from:

• - Failed operations and errors
• User corrections and feedback
• Successful complex task completions
• Periodic review and consolidation

When to Use

1. 1. Error Recovery: When a command or operation fails unexpectedly
2. User Correction: When the user corrects the agent's behavior or output
3. Success Capture: After completing a complex task successfully
4. Learning Review: Periodic review of past learnings to consolidate knowledge

Core Workflow

1. Capture Learning

When something noteworthy happens:

CODEBLOCK0

2. Store Learning

Learnings are stored in memory/learnings.jsonl with:

• - Timestamp
• Type and severity
• Full context and details
• Tags for searchability

3. Retrieve Relevant Learnings

Before starting a task, search past learnings:

• - Match by task type
• Match by tags
• Match by error patterns

4. Apply Learnings

Use retrieved learnings to:

• - Avoid past mistakes
• Apply successful patterns
• Adjust approach based on corrections

Memory Structure

Learnings are stored in JSONL format:

CODEBLOCK1

Learning Types

Severity Levels

• - critical: System-breaking errors, data loss risk
• high: Task-blocking errors, significant issues
• medium: Minor issues, workarounds available
• low: Optimization opportunities, nice-to-know

Commands

Capture a Learning

CODEBLOCK2

Search Learnings

CODEBLOCK3

Review Learnings

CODEBLOCK4

Best Practices

1. 1. Capture Immediately: Record learnings while context is fresh
2. Be Specific: Include full error messages and exact commands
3. Generalize Lessons: Extract principles that apply beyond this instance
4. Tag Thoughtfully: Use consistent tags for easy retrieval
5. Review Regularly: Weekly review helps consolidate knowledge
6. Avoid Duplicates: Check existing learnings before adding new ones

Integration Points

• - Error Handlers: Automatically capture command failures
• User Feedback: Listen for correction patterns in conversation
• Task Completion: Prompt for learning capture after complex tasks
• Heartbeat: Include learning review in periodic checks

Example Scenarios

Scenario 1: Command Failure

CODEBLOCK5

Scenario 2: User Correction

CODEBLOCK6

Scenario 3: Success Pattern

CODEBLOCK7

Safety Rules

• - Never store sensitive data (passwords, API keys, tokens)
• Sanitize error messages that might contain secrets
• Require user approval before storing corrections
• Allow users to delete or edit learnings
• Respect user privacy preferences