self-improving-agent

Self-Improving Agent

This skill enables continuous improvement by tracking errors, lessons learned, and user corrections. Uses persistent database storage.

When to Use

Use this skill proactively in these scenarios:

1. Command or operation fails unexpectedly - A tool execution fails, command returns error
2. User corrects you - User says "no, that's wrong", "not what I meant", or provides corrections
3. Repeated failures - Same type of error occurs multiple times
4. Unknown uncertainty - Not sure if approach is correct, want to verify before proceeding

How It Works

Using the Tool

Use builtin_learning tool with the following operations:

1. Record a Learning (add)

Record what you learned from an error or correction:

operation: "add"
error_type: "shell_command_permission_denied"
context: "Tried to run 'apt-get install' without sudo"
root_cause: "User not in sudoers file"
fix: "Use 'sudo' prefix or check if user has permission"
lesson: "Always check if elevated permissions are needed before system commands"

Parameters:

• operation: "add" or "create"
• error_type: Short identifier (e.g., "shell_permission", "wrong_editor")
• context: What were you trying to do?
• root_cause: Why did it fail?
• fix: What did you do differently?
• lesson: What did you learn? (general principle)
• user_id: User ID (optional, omit for global learning)

2. List Learnings (list)

Get all learnings for a user (or global if no user_id):

operation: "list"
user_id: "123"

3. Get Specific Learning (get)

Retrieve a specific learning by error_type:

operation: "get"
error_type: "shell_command_permission_denied"

Key Principles

• Don't repeat mistakes - Apply learnings to avoid same class of errors
• Be specific - Log concrete details, not vague notes
• Extract patterns - Look for recurring themes
• Acknowledge uncertainty - It's okay to ask for verification when unsure
• Learn from users - Their corrections are valuable feedback