A high-performance memory management system for AI agents utilizing vector search, graph databases, and Karpathy-inspired multi-stage knowledge distillation.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install m-flow-memory
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 m-flow-memory 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 M-Flow Memory Skill?
The M-Flow Memory Skill is a sophisticated memory framework designed for AI agents within the Openclaw Skills ecosystem. It integrates cutting-edge technologies including Cone Graph, LanceDB for vector search, and Kuzu for graph database management to create a robust long-term memory architecture.
By implementing a Karpathy-style 4-stage management workflow, this skill transforms raw session data into structured, searchable knowledge. It allows agents to maintain context over long periods, perform semantic retrieval, and conduct episodic memory searches, making it an essential component for developers building high-autonomy AI systems.
M-Flow Memory Skill Use Cases
- Building a persistent knowledge base for LLMs that evolves over time through automated distillation.
- Enhancing AI agents with episodic memory to recall specific user interactions and session context.
- Implementing structured knowledge retrieval using a mix of lexical (BM25), episodic (vector), and triplet (graph) searches.
- Managing large-scale AI session archives with automated linting and deduplication processes.
How M-Flow Memory Skill Works
- Data Ingestion: Raw session content or specific knowledge points are added to the memory buffer using the add method.
- Knowledge Distillation: The system processes raw data through a 4-stage Karpathy workflow, moving from initial audit to a structured wiki layer.
- Indexing: The memorize function triggers the indexing of all unprocessed content into vector and graph databases.
- Multi-Modal Retrieval: Users or agents query the memory system using various modes including lexical, episodic (vector), or triplet (relational) searches.
- Maintenance: Weekly health checks perform deduplication, merging, and updating of knowledge to ensure database integrity.
M-Flow Memory Skill Setup
To integrate this into your Openclaw Skills environment, ensure you have Ollama running locally and install the specific required dependencies:
# Critical: Use LanceDB version 0.26.0
uv pip install lancedb==0.26.0
# Install M-Flow
pip install m-flow>=0.3.1
Configure your .env file with the following variables:
HF_HUB_OFFLINE=1
LLM_PROVIDER="ollama"
LLM_MODEL="ollama/qwen2.5:14b-instruct-q8_0"
EMBEDDING_PROVIDER="ollama"
VECTOR_DB_PROVIDER=lancedb
GRAPH_DATABASE_PROVIDER=kuzu
M-Flow Memory Skill Data Schema & Taxonomy
The M-Flow Memory Skill organizes data across several layers to ensure high retrieval accuracy within the Openclaw Skills framework:
| Layer | Format | Purpose |
|---|---|---|
| Wiki | source | content | tags | timestamp |
Initial structured storage of raw query results |
| Compile | Knowledge Points | Distilled insights from session transcripts |
| M-Flow | Vector/Graph Embeddings | High-performance search index via LanceDB and Kuzu |
| Archive | Collection | Long-term cold storage of historical data |
M-Flow Memory Skill Advanced Features
- Multi-Agent Sharing: Allows multiple agents to interact with and update a shared memory pool via an MCP server.
- Triple-Mode Search: Support for hybrid retrieval strategies combining BM25, Cone Graph vector search, and Knowledge Graph triplets.
- Automated Distillation: Built-in logic to automatically extract core knowledge points from messy conversational transcripts.
- Custom Knowledge Standards: Support for
KNOWLEDGE-STANDARDS.mdto define how the agent should categorize and validate new information.
SKILL.md
Loading
Related Openclaw Skills
Karpathy Lint Skill
A specialized quality control skill for the Karpathy LLM Knowledge Base designed to perform deduplication, merging, and self-repair of knowledge points.
Chinese Construction Tendering Agent
A professional AI agent designed to manage the full lifecycle of construction bidding and procurement processes within the Chinese legal framework.
Trump AI Persona Skill
A sophisticated AI persona skill that emulates the speech patterns, vocabulary, and specific catchphrases of Donald Trump using a verified database of real quotes.
Karpathy Query to Wiki Feedback
A sophisticated knowledge pipeline that transforms user queries and memory searches into structured wiki entries for continuous LLM learning.
Karpathy Compile Skill - Phase 2
An AI-driven distillation engine that converts raw wiki entries into high-density, structured knowledge points for advanced information retrieval.
China Construction Cost Quick Estimation System
A professional AI agent skill designed for rapid construction investment estimation and cost composition analysis based on the latest GB/T 50500-2024 standards.
Featured*
