An advanced learning framework that captures atomic behaviors and project-specific patterns to turn AI agents into continuously evolving coding partners.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install self-improving-agent-ecc
2. Manual Installation
Copy the skill folder to one of these locations
~/.openclaw/skills/ <project>/skills/ Priority: Workspace > Local > Bundled
3. Prompt Installation
Copy this prompt to OpenClaw to install it automatically.
Help me install self-improving-agent-ecc using Clawhub. If Clawhub is not installed, install it first (npm i -g clawhub).
Prefer to download?
Get the raw skill files in a ZIP archive.
What is Self-Improving Agent?
The Self-Improving Agent is a sophisticated learning system designed to bridge the gap between temporary chat sessions and long-term knowledge retention. By utilizing Openclaw Skills, this agent monitors interactions to capture what it calls instincts—small, atomic learned behaviors that include confidence scoring and project-scope isolation. This ensures that specific React patterns stay in your React projects while universal security best practices are shared across all your development environments.
This system allows AI agents like Claude Code to evolve based on user corrections, discovery of new patterns, and command failures. It moves beyond simple logging by providing a mechanism to cluster these atomic instincts into fully realized skills, slash commands, or specialized sub-agents, creating a truly personalized developer experience.
Self-Improving Agent Use Cases
- Capturing and enforcing project-specific coding conventions and file structures.
- Automating the logging of command failures and identifying systemic fixes.
- Building a persistent memory of user preferences across multiple Openclaw Skills sessions.
- Promoting local project patterns to a global knowledge base after high-confidence verification.
- Using sub-agents to background-process complex architectural learnings.
How Self-Improving Agent Works
- Observation: Deterministic hooks (PreToolUse/PostToolUse) monitor every action and user correction in real-time.
- Instinct Generation: The system logs atomic behaviors as YAML files, assigning confidence scores based on evidence.
- Context Awareness: The agent detects the current project via git hash or environment variables to isolate knowledge.
- Refinement: Confidence levels increase with repeated observations or decrease if the user provides a correction.
- Evolution: Users run the evolve command to cluster related instincts into permanent markdown skills or global automation rules.
Self-Improving Agent Setup
To install via ClawdHub, use the following command:
clawdhub install self-improving-agent
For manual installation into your Openclaw Skills directory:
git clone https://github.com/peterskoett/self-improving-agent.git ~/.openclaw/skills/self-improving-agent
Finally, initialize your workspace learning directory:
mkdir -p ~/.openclaw/workspace/.learnings
Self-Improving Agent Data Schema & Taxonomy
The skill maintains a structured hierarchy within the homunculus directory to separate global and project-specific data.
| Path | Purpose |
|---|---|
identity.json |
Profile and technical level metadata |
projects.json |
Registry mapping project hashes to names and paths |
instincts/personal/ |
Global auto-learned atomic behavioral files (YAML) |
projects/{hash}/ |
Project-isolated observations and instincts |
.learnings/ |
Workspace-level Markdown logs for errors and feature requests |
Self-Improving Agent Advanced Features
- Confidence-Weighted Execution: Behaviors are only auto-applied when reaching high confidence thresholds (e.g., 0.7+).
- Automated Evolution: The evolve command suggests new slash commands and specialized agents based on clustered patterns.
- Deterministic Observation Hooks: Uses 100% reliable shell-level hooks instead of relying on LLM judgment for logging.
- Inter-Session Messaging: Ability to send learnings and transcripts between active AI sessions using Openclaw Skills tools.
- Project Promotion: Logic to auto-promote patterns seen across 2+ distinct projects to the global scope.
SKILL.md
Loading
Related Openclaw Skills
Performance Optimization (Web & LLM Cost)
A comprehensive optimization toolkit for boosting web application speed and minimizing LLM API expenditures through intelligent resource management.
OpenClaw Security Practice Guide
A comprehensive security framework for AI agents with root access, covering lifecycle auditing, permission hardening, and automated disaster recovery.
Openclaw Multi-Agent Telegram System
A technical framework for deploying multiple specialized AI agents in a single Telegram group with a unified workspace and shared memory.
MemOS Memory Recall
A persistent memory skill that enables AI agents to query and recall past interactions and user preferences from MemOS Cloud.
TCM Diagnosis
An advanced AI-driven Traditional Chinese Medicine diagnostic agent providing pattern differentiation and wellness recommendations.
TDD & De-Sloppify Agent Skill
A sophisticated TDD workflow for AI agents featuring a unique de-sloppify pattern to ensure clean, behavior-driven code.
Featured*
