Learning Loop is a structured self-improvement system that provides AI agents with persistent, compounding intelligence through a tiered knowledge promotion framework.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install learning-loop
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 learning-loop 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 Learning Loop?
The Learning Loop is a robust infrastructure designed to stop AI agents from repeating failures by preventing knowledge loss during context compaction. It functions as a persistent memory layer that captures debugging lessons, user feedback, and successes, transforming them into a structured hierarchy of knowledge. By utilizing this skill from the Openclaw Skills collection, developers can ensure their agents get measurably better at their jobs with every single session.
This system implements a sophisticated architecture involving confidence decay and cross-agent knowledge sharing. It moves beyond simple logging by providing a processing layer that detects patterns and promotes them into hard behavioral rules. This ensures that an agent's intelligence compounds over time rather than resetting every time the context window is cleared.
Learning Loop Use Cases
- Capturing root causes and solutions immediately after complex debugging sessions.
- Automatically logging user corrections or praise to refine agent behavior.
- Checking persistent rules and pre-action checklists before performing risky shell operations.
- Maintaining behavioral constraints that survive session compactions and context resets.
- Sharing proven operational rules between different agent instances or teams.
How Learning Loop Works
- Raw interactions, mistakes, and feedback are captured as append-only data in an events log.
- Pattern detection scripts analyze raw events to extract structured lessons with associated confidence scores.
- Lessons that are successfully applied multiple times are promoted to the rules layer for enforcement.
- An exponential confidence decay is applied to rules to flag stale information for manual or automated review.
- The agent loads the active rule set during its boot sequence to ensure immediate compliance with learned constraints.
Learning Loop Setup
To initialize the system within your workspace, run the following command:
bash init.sh /path/to/workspace
After initialization, update your agent's boot instructions to read memory/learning/rules.json and memory/learning/BOOT.md at the start of every session to ensure this Openclaw Skills module is active.
Learning Loop Data Schema & Taxonomy
The system organizes data into a tiered structure to manage the lifecycle of a lesson:
| File | Type | Description |
|---|---|---|
events.jsonl |
Append-only JSON | Raw logs of successes, failures, and feedback. |
lessons.json |
Structured JSON | Intermediate patterns extracted from events. |
rules.json |
Enforced JSON | Hard behavioral constraints with confidence scores. |
metrics.json |
Snapshot | Tracking of improvement trends and save rates. |
pre-action-checklist.md |
Markdown | A human-readable list for risky operation validation. |
Learning Loop Advanced Features
- Cross-agent sharing: Export and import rules between agents with SHA256 integrity verification and conflict detection.
- Ebbinghaus-inspired confidence decay: Automatically flags rules that haven't been validated recently to prevent outdated behavior.
- Multi-layered enforcement: Includes boot sequences, compaction flushes, and heartbeat checks to ensure learning persists.
- Automated self-audit: Generates a health score (A-D) based on 23 distinct checks of the learning loop's integrity.
- Write-Ahead Log (WAL) capture: Ensures critical details are saved even during unexpected context compression.
SKILL.md
Loading
Related Openclaw Skills
Construction PM
A production-tested construction project management toolkit designed for AI agents to track jobs from lead to final payment.
Colony Engagement
A comprehensive toolkit for AI agents to interact with thecolony.cc, facilitating collaborative intelligence through authenticated posting, voting, and engagement tracking.
Auto-Research Skill
An autonomous research agent for OpenClaw that searches, synthesizes, and stores findings with full source tracking and vectorization.
Openclaw YouTube Transcript
A high-performance utility for Openclaw agents to extract text transcripts and subtitles directly from YouTube URLs without the need for local audio processing.
Moltbook Engagement
A production-hardened toolkit for AI agents to engage, post, and track performance on the Moltbook platform with advanced deduplication.
MoltbotDen Engagement
A comprehensive toolkit for AI agents to engage in community discussions, discover compatible peers, and showcase collaborative projects on MoltbotDen.
Featured*
