CozoDB Memory MCP Server

Why Cozo Memory?
LLMs have short-term memory limits. Standard RAG retrieves documents but can't connect facts across time. Cozo Memory gives your AI agent persistent, structured memory – it remembers past conversations, infers relationships, detects contradictions, and explores its knowledge graph – fully on your machine, with optional local LLM integration via Ollama for intelligent actions (cleanup, reflection, summarization, agentic routing).

Most memory stacks combine separate databases: SQLite for facts, Chroma for vector search, NetworkX for graphs. CozoDB replaces all of that with one embedded engine: relational, graph, vector, and full-text search in a single query language, one file, zero sync lag.

Local-first memory for Claude & AI agents with hybrid search, Graph-RAG, and time-travel – runs entirely on your machine. Optional Ollama integration enables LLM-powered actions (cleanup, reflect, summarize, agentic retrieval).

Table of Contents

• Quick Start
• Key Features
• Positioning & Comparison
• Installation
• Integration
• Documentation
• Troubleshooting

Quick Start

Option 1: Install via npm (Recommended)

# Install globally
npm install -g cozo-memory

# Or run directly with npx (no installation needed)
npx cozo-memory

Option 2: Build from Source

git clone https://github.com/tobs-code/cozo-memory
cd cozo-memory
npm install && npm run build
npm run start

Now add the server to your MCP client (e.g. Claude Desktop) – see Integration below.

Key Features

🔍 Hybrid Search - Combines semantic (HNSW), full-text (FTS), and graph signals via Reciprocal Rank Fusion for intelligent retrieval

🧠 Agentic Retrieval - Auto-routing engine analyzes query intent via local LLM to select optimal search strategy (Vector, Graph, or Community)

⏱️ Time-Travel Queries - Version all changes via CozoDB Validity; query any point in history with full audit trails

🎯 GraphRAG-R1-Inspired Adaptive Retrieval - Intelligent system with Progressive Retrieval Attenuation (PRA) and Cost-Aware F1 (CAF) scoring, conceptually inspired by GraphRAG-R1 (Yu et al., WWW 2026) and adapted for CozoDB, that learns from usage

⏳ Temporal Conflict Resolution - Automatic detection and resolution of contradictory observations with semantic analysis and audit preservation

🏠 100% Local - Embeddings via ONNX/Transformers; data stays on your machine. Some advanced features (cleanup, reflect, summarize, agentic search) require an optional Ollama service for local LLM inference — but the core search, CRUD, and graph operations work without any LLM.

🧠 Multi-Hop Reasoning - Logic-aware graph traversal with vector pivots for deep relational reasoning

🗂️ Hierarchical Memory - Multi-level architecture (L0-L3) with intelligent compression and LLM-backed summarization

→ See all features | Version History

Positioning & Comparison

Why CozoDB instead of SQLite + Chroma + NetworkX?

A common first question is: "Why not just combine existing tools?"

The core insight: Most memory stacks bolt vector search onto a graph DB, or graph search onto a vector DB. CozoDB is different: it is a single engine that natively combines relational, graph, vector, and full-text search. That means:

• One query language (Datalog) reaches every dimension.
• No sync lag between separate indexes.
• No ETL bridge between "vector results" and "graph expansion."
• Smaller operational surface: one database file, one process, one dependency chain.

Comparison with other memory solutions

Most "Memory" MCP servers fall into two categories:

1. Simple Knowledge Graphs: CRUD operations on triples, often only text search
2. Pure Vector Stores: Semantic search (RAG), but little understanding of complex relationships

This server fills the gap in between ("Sweet Spot"): A local, database-backed memory engine combining vector, graph, and keyword signals — powered by CozoDB's unified engine rather than a patchwork of separate databases.

The core advantage is Intelligence and Traceability: By combining an Agentic Retrieval Layer with Hierarchical GraphRAG, the system can answer both specific factual questions and broad thematic queries with much higher accuracy than pure vector stores.

Installation

Prerequisites

• Node.js 20+ (recommended)
• RAM: 1.7 GB minimum (for default bge-m3 model)
  • Model download: ~600 MB
  • Runtime memory: ~1.1 GB
  • ⚡ Too heavy? Use EMBEDDING_MODEL=Xenova/all-MiniLM-L6-v2 – only ~400 MB RAM needed (see Embedding Model Options)
• CozoDB native dependency is installed via cozo-node

Optional: Ollama for LLM-powered actions

Some advanced actions use a local LLM via Ollama for intelligent processing. The core server works without Ollama (CRUD, search, graph operations), but the following actions require it:

Setup (if you need these features):

# 1. Install Ollama from https://ollama.ai
# 2. Pull a model (e.g. small + fast for dev):
ollama pull demyagent-4b-i1:Q6_K
# 3. Ollama runs automatically on http://localhost:11434

If Ollama is not running, the affected actions gracefully fall back to non-LLM behavior (where possible) or return a clear error message.

Via npm (Easiest)

# Install globally
npm install -g cozo-memory

# Or use npx without installation
npx cozo-memory

From Source

git clone https://github.com/tobs-code/cozo-memory
cd cozo-memory
npm install
npm run build

Windows Quickstart

npm install
npm run build
npm run start

Notes:

• On first start, @xenova/transformers downloads the embedding model (may take time)
• Embeddings are processed on the CPU

Embedding Model Options

CozoDB Memory supports multiple embedding models via the EMBEDDING_MODEL environment variable:

Configuration Options:

Option 1: Using .env file (Easiest for beginners)

# Copy the example file
cp .env.example .env

# Edit .env and set your preferred model
EMBEDDING_MODEL=Xenova/all-MiniLM-L6-v2

Option 2: MCP Server Config (For Claude Desktop / Kiro)

{
  "mcpServers": {
    "cozo-memory": {
      "command": "npx",
      "args": ["cozo-memory"],
      "env": {
        "EMBEDDING_MODEL": "Xenova/all-MiniLM-L6-v2"
      }
    }
  }
}

Option 3: Command Line

# Use lightweight model for development
EMBEDDING_MODEL=Xenova/all-MiniLM-L6-v2 npm run start

Download Model First (Recommended):

# Set model in .env or via command line, then:
EMBEDDING_MODEL=Xenova/all-MiniLM-L6-v2 npm run download-model

Note: Changing models requires re-embedding existing data. The model is downloaded once on first use.

Integration

Claude Desktop

Using npx (Recommended)

{
  "mcpServers": {
    "cozo-memory": {
      "command": "npx",
      "args": ["cozo-memory"]
    }
  }
}

Using global installation

{
  "mcpServers": {
    "cozo-memory": {
      "command": "cozo-memory"
    }
  }
}

Using local build

{
  "mcpServers": {
    "cozo-memory": {
      "command": "node",
      "args": ["C:/Path/to/cozo-memory/dist/index.js"]
    }
  }
}

Framework Adapters

Official adapters for seamless integration with popular AI frameworks:

🦜 LangChain Adapter

npm install @cozo-memory/langchain @cozo-memory/adapters-core

import { CozoMemoryChatHistory, CozoMemoryRetriever } from '@cozo-memory/langchain';

const chatHistory = new CozoMemoryChatHistory({ sessionName: 'user-123' });
const retriever = new CozoMemoryRetriever({ useGraphRAG: true, graphRAGDepth: 2 });

🦙 LlamaIndex Adapter

npm install @cozo-memory/llamaindex @cozo-memory/adapters-core

import { CozoVectorStore } from '@cozo-memory/llamaindex';

const vectorStore = new CozoVectorStore({ useGraphRAG: true });

Documentation: See adapters/README.md for complete examples and API reference.

CLI & TUI

CLI Tool

Full-featured CLI for all operations:

# System operations
cozo-memory system health
cozo-memory system metrics

# Entity operations
cozo-memory entity create -n "MyEntity" -t "person"
cozo-memory entity get -i <entity-id>

# Search
cozo-memory search query -q "search term" -l 10
cozo-memory search agentic -q "agentic query"

# Graph operations
cozo-memory graph pagerank
cozo-memory graph communities

# Export/Import
cozo-memory export json -o backup.json
cozo-memory import file -i data.json -f cozo

# All commands support -f json or -f pretty for output formatting

See CLI help for complete command reference: cozo-memory --help

TUI (Terminal User Interface)

Interactive TUI with mouse support powered by Python Textual:

# Install Python dependencies (one-time)
pip install textual

# Launch TUI
npm run tui
# or directly:
cozo-memory-tui

TUI Features:

• 🖱️ Full mouse support (click buttons, scroll, select inputs)
• ⌨️ Keyboard shortcuts (q=quit, h=help, r=refresh)
• 📊 Interactive menus for all operations
• 🎨 Rich terminal UI with colors and animations

Architecture Overview

graph TB
    Client[MCP Client<br/>Claude Desktop, etc.]
    Server[MCP Server<br/>FastMCP + Zod Schemas]
    Services[Memory Services]
    Embeddings[Embeddings<br/>ONNX Runtime]
    Search[Hybrid Search<br/>RRF Fusion]
    Cache[Semantic Cache<br/>L1 + L2]
    Inference[Inference Engine<br/>Multi-Strategy]
    DB[(CozoDB SQLite<br/>Relations + Validity<br/>HNSW Indices<br/>Datalog/Graph)]
    
    Client -->|stdio| Server
    Server --> Services
    Services --> Embeddings
    Services --> Search
    Services --> Cache
    Services --> Inference
    Services --> DB
    
    style Client fill:#e1f5ff,color:#000
    style Server fill:#fff4e1,color:#000
    style Services fill:#f0e1ff,color:#000
    style DB fill:#e1ffe1,color:#000

See docs/ARCHITECTURE.md for detailed architecture documentation

MCP Tools Overview

The interface is reduced to 5 consolidated tools:

See docs/API.md for complete API reference with all parameters and examples

Troubleshooting

Common Issues

First Start Takes Long

• The embedding model download takes 30-90 seconds on first start (Transformers loads ~500MB of artifacts)
• This is normal and only happens once
• Subsequent starts are fast (< 2 seconds)

LLM-powered actions require Ollama

• The following actions use a local LLM for intelligent processing: cleanup, reflect, summarize_communities, compact, agentic_search
• Install Ollama from https://ollama.ai
• Pull the desired model: ollama pull demyagent-4b-i1:Q6_K (or your preferred model)
• Without Ollama, these actions fall back to non-LLM behavior or return a clear error
• Core features (CRUD, search, graph, infer) work without any LLM

Windows-Specific

• Embeddings are processed on CPU for maximum compatibility
• RocksDB backend requires Visual C++ Redistributable if using that option

Performance Issues

• First query after restart is slower (cold cache)
• Use health action to check cache hit rates
• Consider RocksDB backend for datasets > 100k entities

See docs/BENCHMARKS.md for performance optimization tips

Documentation

• docs/API.md - Complete MCP tools reference with all parameters and examples
• docs/ARCHITECTURE.md - System architecture, data model, and technical details
• docs/BENCHMARKS.md - Performance metrics, evaluation results, and optimization tips
• docs/FEATURES.md - Detailed feature documentation with usage examples
• docs/USER-PROFILING.md - User preference profiling and personalization
• CHANGELOG.md - Version history and release notes
• CONTRIBUTING.md - Development guidelines

Development

Structure

• src/index.ts: MCP Server + Tool Registration
• src/memory-service.ts: Core business logic
• src/db-service.ts: Database operations
• src/embedding-service.ts: Embedding Pipeline + Cache
• src/hybrid-search.ts: Search Strategies + RRF
• src/inference-engine.ts: Inference Strategies
• src/api_bridge.ts: Express API Bridge (optional)

Scripts

npm run build        # TypeScript Build
npm run dev          # ts-node Start of MCP Server
npm run start        # Starts dist/index.js (stdio)
npm run bridge       # Build + Start of API Bridge
npm run benchmark    # Runs performance tests
npm run eval         # Runs evaluation suite

Contributing

Contributions are welcome! Please see CONTRIBUTING.md for guidelines.

License

Apache 2.0 - See LICENSE for details.

Acknowledgments

Built with:

• CozoDB - Embedded graph database
• ONNX Runtime - Local embedding generation
• Transformers.js - Xenova/bge-m3 model
• FastMCP - MCP server framework

Research foundations:

• GraphRAG-R1 (Yu et al., WWW 2026) - conceptual inspiration for adaptive retrieval
• HopRAG (ACL 2025) - conceptual inspiration for multi-hop reasoning
• T-GRAG (Li et al., 2025) - conceptual inspiration for temporal conflict resolution
• FEEG Framework (Samuel et al., 2026) - conceptual inspiration for query intent classification
• Allan-Poe (arXiv:2511.00855) - conceptual inspiration for dynamic fusion