A persistent memory layer that enables AI agents to record learnings and avoid repeating mistakes across coding sessions.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install mulch-self-improving-agent
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 mulch-self-improving-agent 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 Mulch Self Improver?
Mulch Self Improver is a structured expertise framework designed to solve the "goldfish memory" problem in AI coding agents. While agents typically start every session with a blank slate, this skill allows them to document failures, successful patterns, and architectural decisions in a git-tracked directory.
By integrating this into Openclaw Skills, developers ensure their agents ground themselves in project-specific expertise. This passive layer does not rely on an external LLM for storage, instead using a local, queryable JSONL format that lives alongside your code. This leads to more consistent output, improved developer experience, and significantly reduced hallucinations by grounding the agent in project-specific history.
Mulch Self Improver Use Cases
- When a command, tool execution, or API call fails and requires a specific resolution.
- When the user corrects the agent's logic, approach, or coding style.
- When a new architectural decision or tech stack convention is established for the project.
- When a more efficient coding pattern or best practice is discovered during development.
- To ground a new agent or session in existing project expertise immediately upon startup.
How Mulch Self Improver Works
- Start the session by running mulch prime to load existing project expertise and domain-specific records into the agent's context.
- During the development workflow, the agent uses mulch record to capture failures, conventions, or decisions as they occur.
- The skill utilizes an append-only JSONL format stored in the .mulch/ directory, ensuring data is git-trackable and safe for multi-agent use.
- Before concluding a session, the agent reviews its actions and records any final insights to compound knowledge for future tasks.
- High-value patterns that prove successful over time are promoted to project-wide documentation like CLAUDE.md or AGENTS.md using the onboarding utility.
Mulch Self Improver Setup
Install the CLI tool and initialize it within your project directory to get started with Openclaw Skills.
# Install the CLI globally (optional)
npm install -g mulch-cli
# Initialize Mulch in your current project
mulch init
# Install the skill via OpenClaw
clawdhub install self-improving-agent
# Enable the reminder hook for your workspace
cp -r hooks/openclaw ~/.openclaw/hooks/self-improvement
openclaw hooks enable self-improvement
Mulch Self Improver Data Schema & Taxonomy
Data is organized within the .mulch/ directory using append-only JSONL files, ensuring atomic writes and git compatibility. This structure allows Openclaw Skills to maintain a clean history of technical evolution.
| Record Type | Required Fields | Primary Use Case |
|---|---|---|
| failure | description, resolution | Documenting what went wrong and the specific fix |
| convention | content | Project-specific rules (e.g., "Use pnpm instead of npm") |
| decision | title, rationale | Architecture choices and tech stack tracking |
| pattern | name, description | Reusable code structures or logic flows |
| guide | name, description | Step-by-step procedures for complex tasks |
| reference | name, description | Mapping key files, endpoints, or external resources |
Mulch Self Improver Advanced Features
- Auto-Detection: Automatically detects errors, command failures, and user corrections to prompt the agent for a record entry.
- Multi-Agent Safety: Includes advisory file locking and atomic writes to prevent data corruption when multiple agents work simultaneously.
- Pre-loaded Domains: Comes with 24 preset domains (api, database, security, etc.) to categorize expertise immediately.
- Promotion Workflow: Use mulch onboard to generate snippets for project-level rule files like AGENTS.md or SOUL.md.
- Notification System: Optional integration to notify users via Telegram whenever a new learning is recorded.
- Cross-Provider Hooks: Native support for setting up hooks in Cursor, Claude, Windsurf, Aider, and other popular IDEs.
SKILL.md
Loading
Related Openclaw Skills
FACEPALM
An automated troubleshooting tool that uses Codex 5.3 to analyze console logs and chat history for instant issue resolution.
Systematic Debugging
A rigorous five-phase debugging protocol that enforces root cause investigation before any code fixes are attempted.
Humanizer
A specialized editing agent for Openclaw Skills that identifies and eliminates 24 distinct AI-generated writing patterns to create natural, soulful text.
Content Pipeline Orchestrator
A 4-stage content orchestration system that automates research, ideation, drafting, and queue management for creators.
Markdown to Word Converter
A Python-powered utility to transform Markdown content into formatted Word documents with support for templates and complex elements.
Mermaid Workflow Skill
A technical automation suite for generating, rendering, and embedding Mermaid diagrams into Markdown documentation via a structured script-based workflow.
Featured*
