DOCSEARCHbyAlgolia
Submit Docs

MetaClaw Documentation

repository·main·Indexed 25 days ago

https://github.com/aiming-lab/metaclaw

An agentic framework enabling LLMs to meta-learn and evolve from real-world conversations. MetaClaw acts as an intelligent proxy for skill injection, long-term memory management, and reinforcement learning (RL) without requiring a local GPU cluster. The aiming-metaclaw package (v0.4.1) supports OpenClaw skill injection and RL training with one-click deployment.

Tokens
65.6K
Snippets
139
Records
438
Agent score
79%

What's inside MetaClaw

  1. Overview of MetaClaw

    main

    MetaClaw is an agent framework designed to meta-learn and evolve through real-world conversations. It acts as a proxy that intercepts interactions from personal agents (such as OpenClaw, CoPaw, IronClaw, etc.), injects relevant skills at each turn, and uses accumulated experience to perform continuous learning.

    Key features include:

    • One-click deployment: Automated setup for personal agents.
    • Three operating modes: skills_only (proxying and skill injection), rl (skills + immediate RL training), and auto (skills + RL with a smart scheduler that runs updates during idle/sleep windows).
    • Long-term memory: Persists facts and preferences across sessions.
    • Asynchronous design: Decouples serving, reward modeling, and training so the agent remains responsive during optimization.

  2. Overview of MetaClaw architecture and capabilities

    main

    MetaClaw is a meta-learning agent framework that evolves through real-world interactions. It places your model behind an OpenAI-compatible proxy (or an Anthropic-compatible /v1/messages endpoint for agents like NanoClaw) to intercept interactions from supported agents like OpenClaw, NanoClaw, and NemoClaw.

    Key capabilities include:

    • Skill Injection: Automatically injects relevant skills into every turn and summarizes them after sessions.
    • Long-term Memory: Maintains facts, preferences, and project history across sessions to provide contextually relevant injection.
    • Asynchronous Design: The service, reward modeling, and training processes are fully decoupled. The agent remains responsive while evaluation and optimization run in the background.
    • No GPU Required: Works with any OpenAI-compatible LLM API and uses a Tinker-compatible backend for cloud-based LoRA training.

  4. Overview of Day 12 Data Processing Tasks

    main

    The Day 12 tasks involve processing specific datasets to generate analytical reports. The required tasks are:

    1. Analyze monthly API call statistics.
    2. Generate a user activity report.
    3. Compile exception event statistics.
    4. Output a monthly summary dataset.

    All generated outputs must be in JSON format and must include proper naming conventions and complete metadata fields.

  5. Overview of MetaClaw Architecture and Functionality

    main

    MetaClaw is an agent designed for continuous meta-learning and evolution through real-world interactions. Instead of relying solely on offline training, MetaClaw transforms live conversations into learning signals to improve during deployment.

    Key Capabilities

    • Meta-Learning: Intercepts interactions with supported agents (OpenClaw, NanoClaw, NemoClaw, etc.) via an OpenAI-compatible proxy.
    • Skill Injection: Injects relevant skills into every turn of a conversation.
    • Asynchronous RL: When RL is active, weight updates are scheduled during idle periods (e.g., sleep time or inactive periods) to prevent disruption during active use.
    • No GPU Required: Works directly with any OpenAI-compatible LLM API and uses Tinker-compatible backends for cloud-based LoRA training.
    • Memory Layer: Features a contexture layer that maintains memory across sessions, automatically retrieving relevant facts, preferences, and project history.

  6. Overview of Day 18 Code Engineering Tasks

    main

    The Day 18 task set involves performing critical modifications to the production system. The tasks include:

    1. Update Production Configuration: Adjusting timeout and logging settings in prod_config.py.
    2. Refactor Log Handler Module: Converting the synchronous log writing in log_handler.py to an asynchronous queue-based pattern to prevent main thread blocking.
    3. Fix API Route Bug: Correcting route path inconsistencies in api_routes.py.
    4. Update Dependency Versions: Upgrading library versions in requirements.txt to newer stable releases.

  7. Review Day 09 - Sprint 8 Project Management tasks

    main

    The Day 09 project management workspace for Sprint 8 involves the following tasks:

    1. Sprint mid-point progress review
    2. Update milestone status
    3. Prepare next week's staffing schedule
    4. Generate risk assessment report

    Required materials for these tasks include:

    • sprint8_progress_raw.txt: Raw progress data from the team.
    • resource_availability.txt: Next week's team availability.

  9. Day 24 Code Engineering Tasks Overview

    main

    This document outlines the specific code engineering tasks scheduled for April 16, 2026. Tasks include developing Python utilities for data validation and report generation, updating application configurations, creating JSON test fixtures for API endpoints (/projects and /tasks), and setting up integration test configurations.

    All deliverables must adhere to the following project requirements:

    • P1: Use ISO 8601 timestamps.
    • P3: Ensure metadata is present in all files.
    • P5: Include done.log entries.

  10. Understand MetaClaw operating modes

    main

    MetaClaw operates in three distinct modes depending on your requirements for Reinforcement Learning (RL) and scheduling:

    ModeDefaultDescription
    skills_onlyProvides your LLM API via a proxy. Injects Skills and performs auto-summarization after sessions. Does not require GPU or Tinker.
    rlIncludes Skills + RL training (GRPO). Training starts immediately as batches fill. Supports optional OPD for teacher distillation.
    auto✅Includes Skills + RL + a smart scheduler. RL weight updates are scheduled to run only during sleep, idle, or meeting windows.

  12. Use MetaClaw memory slash commands and AI tools

    main

    MetaClaw provides several ways for users and LLMs to interact with the memory system.

    Slash Commands (User Interface)

    • /remember <text>: Save information to long-term memory.
    • /recall <query>: Search memories by query.
    • /memory-status: Show memory system health.

    AI Tools (LLM Interface)

    The plugin registers 4 tools that the LLM can call directly:

    • metaclaw_memory_search: Search long-term memories by keyword or semantic query.
    • metaclaw_memory_store: Store a new memory explicitly.
    • metaclaw_memory_forget: Archive a specific memory by ID.
    • metaclaw_memory_status: Get memory system health and statistics.