mem0-open-mcp 0.2.0

Open-source MCP server for mem0 - local LLMs, self-hosted, Docker-free

mem0-open-mcp

Open-source MCP server for mem0 — local LLMs, self-hosted, Docker-free.

Created because the official mem0-mcp configuration wasn't working properly for my setup.

Features

• Local LLMs: Ollama (recommended), LMStudio*, or any OpenAI-compatible API
• Self-hosted: Your data stays on your infrastructure
• Docker-free: Simple pip install + CLI
• Flexible: YAML config with environment variable support
• Multiple Vector Stores: Qdrant, Chroma, Pinecone, and more

*LMStudio requires JSON mode compatible models

Quick Start

Installation

pip install mem0-open-mcp

Or install from source:

git clone https://github.com/wonseoko/mem0-open-mcp.git
cd mem0-open-mcp
pip install -e .

Usage

# Create default config
mem0-open-mcp init

# Interactive configuration wizard
mem0-open-mcp configure

# Test configuration (recommended for initial setup)
mem0-open-mcp test

# Start the server
mem0-open-mcp serve

# With options
mem0-open-mcp serve --port 8765 --user-id alice

The test command verifies your configuration without starting the server:

• Checks Vector Store, LLM, and Embedder connections
• Performs actual memory add/search operations
• Cleans up test data automatically

Configuration

Create mem0-open-mcp.yaml:

server:
  host: "0.0.0.0"
  port: 8765
  user_id: "default"

llm:
  provider: "ollama"
  config:
    model: "llama3.2"
    base_url: "http://localhost:11434"

embedder:
  provider: "ollama"
  config:
    model: "nomic-embed-text"
    base_url: "http://localhost:11434"
    embedding_dims: 768

vector_store:
  provider: "qdrant"
  config:
    collection_name: "mem0_memories"
    host: "localhost"
    port: 6333
    embedding_model_dims: 768

With LMStudio

⚠️ Note: LMStudio requires a model that supports response_format: json_object. mem0 uses structured JSON output for memory extraction. If you get response_format errors, use Ollama instead or select a model with JSON mode support in LMStudio.

llm:
  provider: "openai"
  config:
    model: "your-model-name"
    base_url: "http://localhost:1234/v1"

embedder:
  provider: "openai"
  config:
    model: "your-embedding-model"
    base_url: "http://localhost:1234/v1"

MCP Integration

Connect your MCP client to:

http://localhost:8765/mcp/<client-name>/sse/<user-id>

Claude Desktop

{
  "mcpServers": {
    "mem0": {
      "url": "http://localhost:8765/mcp/claude/sse/default"
    }
  }
}

Available MCP Tools

Tool	Description
add_memories	Store new memories from text
search_memory	Search memories by query
list_memories	List all user memories
get_memory	Get a specific memory by ID
delete_memories	Delete memories by IDs
delete_all_memories	Delete all user memories

API Endpoints

Endpoint	Method	Description
/health	GET	Health check
/api/v1/status	GET	Server status
/api/v1/config	GET/PUT	Configuration
/api/v1/memories	GET/POST/DELETE	Memory operations
/api/v1/memories/search	POST	Search memories

Requirements

• Python 3.10+
• Vector store (Qdrant recommended)
• LLM server (Ollama, LMStudio, etc.)

Graph Store (Experimental)

Graph store enables knowledge graph capabilities for relationship extraction between entities.

Configuration

graph_store:
  provider: "neo4j"
  config:
    url: "bolt://localhost:7687"
    username: "neo4j"
    password: "your-password"

Installation

pip install mem0-open-mcp[neo4j]
# or
pip install mem0-open-mcp[kuzu]

Limitations

⚠️ Important: Graph store requires LLMs with proper tool calling support.

• OpenAI models: Full support (recommended for graph store)
• Ollama models: Limited support - most models (llama3.2, llama3.1) do not follow tool schemas accurately, resulting in empty graph relations

If you need graph capabilities with local LLMs, consider using the graph_store.llm setting to specify a different LLM provider for graph operations only.

# Example: Use OpenAI for graph, Ollama for everything else
llm:
  provider: "ollama"
  config:
    model: "llama3.2"

graph_store:
  provider: "neo4j"
  config:
    url: "bolt://localhost:7687"
    username: "neo4j"
    password: "password"
  llm:
    provider: "openai"
    config:
      model: "gpt-4o-mini"

License

Apache 2.0