memory-hub-mcp 1.9.0

Local Memory Hub MCP Server with stdio transport for ZenCoder and MCP clients

Memory Hub MCP Server (UV/UVX)

A local memory hub for AI agents with MCP integration, designed for ZenCoder and other MCP clients using stdio transport.

Quick Start with UVX

Installation & Usage

# Install and run directly with uvx
uvx memory-hub-mcp

# Or install locally first
uv pip install memory-hub-mcp
memory-hub-mcp

For ZenCoder Integration

In ZenCoder's custom MCP server configuration, you must now provide the URLs for the dependent services (Qdrant and LM Studio).

Command: uvx

Arguments:

[
    "memory-hub-mcp",
    "--qdrant-url",
    "http://<ip_address_of_qdrant>:6333",
    "--lm-studio-url",
    "http://<ip_address_of_lm_studio>:1234/v1"
]

Note: Replace <ip_address_...> with the actual IP addresses where your services are running. If they are on the same machine, the IP will be the same for both.

Development Setup

# Clone and setup
git clone <your-repo>
cd memory-hub
uv venv
source .venv/bin/activate
uv pip install -e .

# Run in development
memory-hub-mcp --log-level DEBUG --qdrant-url http://localhost:6333 --lm-studio-url http://localhost:1234/v1

Publishing to PyPI

To publish a new version of the package to PyPI:

1. Update the Version: Increment the version number in pyproject.toml. PyPI does not allow re-uploading the same version.

   # pyproject.toml
   [project]
   name = "memory-hub-mcp"
   version = "0.1.2" # Increment this
   

2. Clean and Rebuild: Remove old builds and create the new distributions.

   rm -rf dist/
   uv build
   

3. Publish with an API Token:

   The recommended way to publish is to use a PyPI API token. You can provide it directly to the command via an environment variable for security.

   # Replace <your_pypi_token> with your actual token
   UV_PUBLISH_TOKEN=<your_pypi_token> uv publish dist/*
   

Available Tools

• add_memory: Store content with hierarchical metadata (app_id, project_id, ticket_id)
• search_memories: Semantic search with keyword enhancement and LLM synthesis
• get_project_memories: Retrieve ALL memories for a specific app_id/project_id without search queries
• update_memory: Update existing memories with automatic version incrementing
• get_recent_memories: Retrieve memories from the last N hours (perfect for resuming work)
• list_app_ids: List all application IDs
• list_project_ids: List all project IDs
• list_ticket_ids: List all ticket IDs
• list_memory_types: List memory types currently in use (with counts and metadata)
• get_memory_type_guide: Get the recommended memory type conventions
• health_check: Server health status

Configuration

The server expects:

• Qdrant: Vector database running (see docker-compose.yml)
• LM Studio: For embeddings and chat completions
• Environment: Standard .env configuration

Key File & Directory Locations

• pyproject.toml: Defines project metadata, dependencies, and the memory-hub-mcp script entry point.
• src/memory_hub/: The main Python package source code.
• src/memory_hub/cli.py: The command-line interface logic that launches the server.
• src/memory_hub/mcp_server.py: Core stdio server implementation and tool registration.
• src/memory_hub/core/handlers/: Contains the implementation for each MCP tool (e.g., add_memory, search_memories).
• src/memory_hub/core/services.py: Handles communication with external services like Qdrant and LM Studio.
• src/memory_hub/core/models.py: Pydantic models defining the data structures used throughout the application.
• docker-compose.yml: Defines the Qdrant service dependency.

Architecture

• stdio transport: Direct MCP protocol communication
• No HTTP dependencies: Lightweight, focused on MCP clients
• Hierarchical memory: Flexible app/project/ticket organization
• Hybrid search: Vector similarity + keyword matching + LLM synthesis
• Version management: Automatic versioning for memory updates
• Time-based retrieval: Query recent memories by hours

Agent Usage Guide

Saving Agent Progress

# Save initial work
add_memory(
    content="Implemented user authentication with JWT tokens...",
    metadata={
        "app_id": "eatzos",
        "project_id": "next",
        "type": "feature_implementation"
    }
)

# Update existing memory
update_memory(
    app_id="eatzos",
    project_id="next", 
    memory_type="feature_implementation",
    new_content="Completed authentication with JWT tokens and added refresh token logic..."
)

Resuming Agent Work

# Get ALL context for a project (no search guessing!)
get_project_memories(
    app_id="eatzos",
    project_id="next",
    limit=50
)

# See what changed recently
get_recent_memories(
    app_id="eatzos",
    hours=24,
    include_summary=True
)

Differences from HTTP Version

This UV/UVX version:

• ✅ Uses stdio transport (ZenCoder compatible)
• ✅ No FastAPI dependencies
• ✅ Lightweight packaging
• ✅ Direct MCP protocol
• ❌ No web interface
• ❌ No HTTP endpoints