A sophisticated agent learning system that filters, scores, and promotes corrections and patterns into permanent project memory through a quality-gated pipeline.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install midos-self-improver
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 midos-self-improver 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 midos-self-improver?
The midos-self-improver skill provides a robust framework for AI agents to evolve based on real-world interactions. Unlike typical systems that suffer from knowledge base pollution, this Openclaw Skills entry uses a multi-stage pipeline consisting of capture, quality gating, staging, scoring, and promotion. It ensures that only recurring, high-impact insights become part of the permanent project memory.
By implementing this skill, developers can transform transient session data into durable rules. This prevents the agent from repeating the same mistakes and allows it to adopt best practices specific to your codebase. As a core part of the Openclaw Skills ecosystem, it bridges the gap between raw tool outputs and refined, actionable project intelligence.
midos-self-improver Use Cases
- Automatically capturing and applying user corrections to prevent repetitive git or syntax mistakes.
- Analyzing recurring tool errors to generate preventative rules in project documentation.
- Identifying high-value coding patterns across multiple domains for promotion to CLAUDE.md.
- Filtering out low-value noise and docstring-only edits to keep the knowledge base lean and relevant.
How midos-self-improver Works
- Detection triggers monitor the agent session for five specific types: corrections, errors, gaps, best practices, and patterns.
- Potential learnings are sent through a deterministic Quality Gate that performs SHA-256 deduplication and a decision check.
- Validated entries are moved to a staging area where they are tracked for recurrence and freshness over a 14-day half-life.
- A 4-axis scoring system calculates a composite value based on recurrence, freshness, specificity, and impact.
- Learnings that cross the 0.7 composite score threshold are promoted to the permanent knowledge base or project rule files.
midos-self-improver Setup
To integrate this capability into your Openclaw Skills environment, you can use the standalone mode by adding the protocol to your project instructions. Begin by preparing the directory structure:
mkdir -p .learnings/entries .patterns .knowledge
Next, define the promotion thresholds in your agent's system prompt to ensure it tracks corrections and errors. For high-performance environments, you can invoke the pattern harvester via Python:
python -c "from hooks.pattern_harvester import assess_pattern_value; assess_pattern_value()"
midos-self-improver Data Schema & Taxonomy
The skill manages data across a tiered directory structure to maintain organization and clarity. Openclaw Skills users can inspect the following hierarchy:
| Directory | Data Type | Retention Policy |
|---|---|---|
.learnings/ |
Raw JSON entries | 30-day window |
.patterns/ |
Markdown staging files | Until promoted or pruned |
.knowledge/ |
Permanent promoted rules | Persistent |
JSON Entry Metadata:
id: Unique hash based on normalized content.type: One of 5 trigger categories.scores: A composite of recurrence (0.35), freshness (0.25), specificity (0.20), and impact (0.20).status: Current lifecycle state (staging, promoted, or pruned).
midos-self-improver Advanced Features
- Multi-tier promotion architecture moving from entries to staging to chunks and finally to rules.
- SHA-256 and trigram similarity checking to prevent duplicate learnings in your Openclaw Skills database.
- Deterministic noise rejection that ignores trivial maintenance edits and non-decision logs.
- Automated pruning of low-value data (composite score < 0.3) to prevent memory bloating.
- Seamless integration with MidOS ecosystem features like GEPA coherence scoring and L2R reranking.
SKILL.md
Loading
Related Openclaw Skills
Next.js App Router Production Patterns
A production-grade implementation guide for Next.js 15 and 16 focusing on async APIs, caching changes, and Turbopack performance.
MidOS MCP
MidOS is a comprehensive knowledge middleware layer that provides AI agents with long-term memory, hybrid search, and multi-step planning capabilities.
Go Language Patterns & Best Practices
A definitive collection of production-ready Go 1.24+ patterns covering concurrency, web services, database access, and testing.
FastAPI Production Patterns
Accelerate backend development with production-ready FastAPI patterns for Pydantic v2, async workflows, and secure deployments.
Rust Systems Programming Patterns
A production-grade collection of Rust patterns for async concurrency, web services, and systems integration using the 2024 edition.
TypeScript Mastery
Master advanced TypeScript patterns like branded types and discriminated unions to build industrial-grade, type-safe applications.
Featured*
