self-improvement

What This Skill Does

The self-improvement skill for OpenClaw is an industrial-grade, closed-loop system designed to manage how an AI agent evolves. Instead of letting an agent haphazardly update its core instructions based on emotional impulses or one-off bugs, this skill implements the 'Rule of 3'. By tracking errors, corrections, and requests in a structured repository (~/.openclaw/workspace/memory/core), the agent waits for a pattern to emerge before promoting a temporary learning to a permanent fixture in your SOUL.md, AGENTS.md, or TOOLS.md files. This ensures your configuration remains lean, accurate, and noise-free.

Installation

Installation is streamlined through the ClawdHub CLI. Ensure your environment is initialized for OpenClaw before executing the following command:

clawhub install self-improving-agent

Once installed, verify the directory structure is created in ~/.openclaw/workspace/memory/core containing learning.md, error.md, and features.md. These files serve as the primary audit trail for the agent's growth cycle.

Use Cases

This skill is indispensable for long-running development projects where context bloat is a major risk. Use it when:

• An API call consistently fails and requires a specific workaround.
• You repeatedly correct the agent on its tone or output format.
• You frequently request a tool feature that the agent currently lacks.
• You observe a recurring logical error in complex codebase refactoring tasks.
• You wish to standardize your workflow without manually editing configuration files after every single interaction.

Example Prompts

1. "I noticed you keep failing to run the git commit command because of the GPG signing error; please log this as a recurring error in your self-improvement cycle."
2. "Actually, for our current project, I prefer that you always provide a summary before showing the code diff. Add this as a learning and track it for potential promotion to the SOUL.md file."
3. "Review your pending learnings and let me know if any issues have hit the Rule of 3 threshold so we can update our agent workflow."

Tips & Limitations

• Be Patient: The Rule of 3 is a feature, not a bug. Do not force-promote learnings that haven't occurred repeatedly, as this degrades the quality of your system instructions.
• Human Oversight: Always run the promote-review.sh script to audit what the agent intends to write into your core files. Never grant the agent full autonomous write access to SOUL.md without verifying the logic first.
• Organization: Keep your logs clean. If a specific error is resolved by a system update, mark it as 'resolved' in error.md to prevent it from being incorrectly flagged for promotion.