An advanced meta-cognition layer that enables AI agents to evaluate their own work, learn from corrections, and organize compounding knowledge automatically.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install self-improving-1-2-10
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-1-2-10 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 (Proactive Self-Reflection)?
The Self-Improving Agent is a sophisticated framework designed to transform a standard AI agent into a learning entity that evolves with every interaction. By implementing a proactive self-reflection loop, the agent evaluates its performance after significant tasks, identifying mistakes and optimizing its future behavior. This skill is a cornerstone for users of Openclaw Skills who require an agent that doesn't just follow instructions, but actually improves its internal logic and stylistic adherence over time.
At its core, the skill manages a tiered memory architecture located in the local file system. This ensures that the most relevant instructions (HOT memory) are always in context, while older or project-specific patterns (WARM/COLD memory) are swapped in only when needed. This approach prevents context window bloat while ensuring that hard-won lessons are never lost, allowing your agent's capabilities to compound indefinitely.
Self-Improving Agent (Proactive Self-Reflection) Use Cases
- When an agent receives explicit corrections or stylistic feedback from a user.
- After completing complex, multi-step workflows where quality assessment is critical.
- Identifying and automating recurring patterns in code, writing, or communication.
- Maintaining project-specific knowledge bases that persist across multiple sessions.
- Improving agent reliability by logging and correcting recurring failure modes.
How Self-Improving Agent (Proactive Self-Reflection) Works
- Monitoring and Detection: The agent actively listens for correction triggers (e.g., "Actually, it should be...") and preference signals.
- Self-Reflection Cycle: Upon task completion, the agent performs a self-critique to compare its output against the user's intent and logs improvements.
- Tiered Data Entry: New lessons are initially logged in corrections.md or specific project files to keep the primary context clean.
- Pattern Promotion: If a specific lesson or pattern is successfully applied three times within seven days, it is promoted to HOT memory (memory.md).
- Automated Maintenance: The skill automatically demotes unused patterns to WARM or COLD storage based on a 30/90-day inactivity threshold to ensure optimal performance.
Self-Improving Agent (Proactive Self-Reflection) Setup
To initialize the self-reflection framework, install the skill via your terminal:
clawhub install self-improving
If the directory ~/self-improving/ does not exist on your system, the agent will guide you through the initial configuration or you can manually initialize it to begin the automated learning process. Ensure the agent has read/write permissions for its dedicated memory folder.
Self-Improving Agent (Proactive Self-Reflection) Data Schema & Taxonomy
The skill organizes its compounding knowledge using a tiered file structure to maximize efficiency for Openclaw Skills:
| Tier | Storage Location | Size Limit | Purpose |
|---|---|---|---|
| HOT | memory.md |
≤100 lines | Global rules and high-frequency patterns always loaded. |
| WARM | projects/, domains/ |
≤200 lines/file | Context-specific knowledge loaded on demand. |
| COLD | archive/ |
Unlimited | Historical patterns and decayed rules for manual retrieval. |
| LOGS | corrections.md |
Last 50 items | Raw log of recent user feedback and self-reflections. |
Self-Improving Agent (Proactive Self-Reflection) Advanced Features
- Namespace Isolation: Keeps project-specific learnings contained within
projects/{name}.mdto prevent cross-context confusion. - Conflict Resolution Engine: Automatically prioritizes the most specific or most recent rules when instructions contradict (Project > Domain > Global).
- Automated Compaction: Merges similar feedback entries into unified rules to stay within file size limits without losing data.
- Memory Stats Reporting: Provides detailed analytics on how many patterns have been promoted, demoted, or archived.
- Transparent Citations: The agent explicitly cites its memory sources (e.g., "Using X from projects/ui-dev.md") so you always know why it made a specific choice.
SKILL.md
Loading
Related Openclaw Skills
Skill Vetter
A security-first vetting protocol designed to identify red flags and malicious code in AI agent skills before they are installed.
OpenClaw Configuration Manager
A comprehensive technical reference and management tool for the openclaw.json configuration file that powers the AI agent gateway.
Skill Creator
The Skill Creator provides a standardized framework and CLI tools for building, validating, and packaging modular extensions that enhance AI agent capabilities.
ClawConnect
A unified API bridge connecting AI agents to Twitter, Gmail, Slack, and Discord for seamless cross-platform communication.
Agent Reach
A comprehensive configuration and installation toolset that gives AI agents direct access to over 13 major social and web platforms.
Yida Dev Assistant
A specialized tool for DingTalk Yida low-code development, facilitating form creation, JS action scripting, and workflow automation.
Featured*
