Medha MCP

Git-backed AI Memory System

Medha is a Model Context Protocol (MCP) server that provides persistent, git-versioned memory storage for LLM applications. It combines the power of Git's version control with graph-based memory associations, enabling LLMs to maintain long-term memory.

Features

• 🔐 Flexible Authentication: Local mode (development) or SAML 2.0 (production with DUO/Okta)
• 📝 Git-Backed Storage: Every memory change is a git commit with full history
• 🕸️ Graph Associations: Link memories with typed relationships
• 🔍 Powerful Search: Search by tags, dates, content, and associations
• 🧠 Semantic Search: Optional AI-powered vector search with OpenAI embeddings
• 📊 Knowledge Graphs: Traverse memory associations with N-hop queries
• 🔄 Auto-Sync: Hourly synchronization to GitHub with PAT authentication
• 💾 Dual Storage: Git repository (primary) + per-user SQL database (index)
• 🗑️ Soft Delete: Archive memories while preserving complete history
• 🏢 Multi-Database: SQLite for development / local, PostgreSQL for system database
• 🔒 Optimistic Locking: Safe concurrent access from multiple AI agents

Architecture

Requirements

• Go 1.24+ - For building from source
• Git 2.x+ - Required for version-controlled memory storage (git must be in your PATH)
• GCC/Build Tools - Required for CGO (sqlite-vec vector search)
  • macOS: xcode-select --install
  • Linux: sudo apt install build-essential
  • Windows: Use Docker (recommended) or MinGW

Quick Start

1. Build

make setup    # Run initial setup
make deps     # Install dependencies
make build    # Build the binary

2. MCP Client Integration

Option 1: Native Binary

Add to your MCP client config:

Cursor (~/.cursor/mcp_settings.json):

{
  "mcpServers": {
    "medha": {
      "command": "/path/to/bin/medha",
      "env": {
        "ENCRYPTION_KEY": "your-32-char-encryption-key-here"
      }
    }
  }
}

Claude Desktop (~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "medha": {
      "command": "/path/to/bin/medha",
      "args": [],
      "env": {
        "ENCRYPTION_KEY": "your-32-char-encryption-key-here"
      }
    }
  }
}

Option 2: Go Run

Run Medha directly using go run without installing - similar to uvx for Python or npx for Node.js.

Prerequisites: Install Go 1.24+ go.dev/dl and Git in your PATH:

# macOS (Homebrew)
brew install go git

# Ubuntu/Debian
sudo apt update && sudo apt install golang-go git

# Windows (Chocolatey)
choco install golang git

MCP Client Config:

{
  "mcpServers": {
    "medha": {
      "command": "go",
      "args": ["run", "github.com/tejzpr/medha-mcp/cmd/server@latest"],
      "env": {
        "ENCRYPTION_KEY": "your-32-char-encryption-key-here"
      }
    }
  }
}

Option 3: Docker Container

Configure your MCP client to use the Docker Hub image (no build required).

Simple mode (single user):

Uses whoami inside container, which returns medha. All memories are stored under the medha user.

{
  "mcpServers": {
    "medha": {
      "command": "docker",
      "args": [
        "run", "-i", "--rm",
        "-v", "/Users/yourname/.medha:/home/medha/.medha",
        "-e", "ENCRYPTION_KEY=your-32-char-encryption-key-here",
        "tejzpr/medha-mcp"
      ]
    }
  }
}

Multi-user mode (with --with-accessinguser):

Uses ACCESSING_USER env var for user identity. Useful when multiple users share the same Medha instance or when integrating with authenticated systems.

{
  "mcpServers": {
    "medha": {
      "command": "docker",
      "args": [
        "run", "-i", "--rm",
        "-v", "/Users/yourname/.medha:/home/medha/.medha",
        "-e", "ENCRYPTION_KEY=your-32-char-encryption-key-here",
        "-e", "ACCESSING_USER=yourname",
        "tejzpr/medha-mcp",
        "--with-accessinguser"
      ]
    }
  }
}

Important:

• Replace /Users/yourname with your actual home directory path (e.g., /Users/john on macOS, /home/john on Linux)
• Replace yourname in ACCESSING_USER with your actual username
• MCP clients do not expand $HOME or ~ in JSON configs - use absolute paths
• The .medha mount persists your database and git repositories
• The container runs as user medha (uid 1000), data is stored at /home/medha/.medha

Build locally (optional):

docker build -t medha-mcp .

Then use medha-mcp instead of tejzpr/medha-mcp in your config.

Required mounts:

Mount	Purpose
/Users/yourname/.medha:/home/medha/.medha	Persists database, config, and git repositories

Environment variables:

Variable	Purpose
ENCRYPTION_KEY	32-character key for encrypting PAT tokens
ACCESSING_USER	Username for memory isolation (only with --with-accessinguser)

Flags explained:

• -i = Interactive mode (required for stdio transport)
• --rm = Remove container when done
• --with-accessinguser = Use ACCESSING_USER env var instead of whoami for user identity

3. Run Modes

stdio mode (default): For MCP client integration

./bin/medha                      # Uses system user (whoami)
./bin/medha --with-accessinguser # Uses ACCESSING_USER env var

• Communicates via stdin/stdout
• Automatically authenticates user
• No web interface needed

HTTP mode: For web interface

./bin/medha --http              # Default port 8080
./bin/medha --http --port=9000  # Custom port

• Provides web authentication at http://localhost:8080/auth
• Supports SAML 2.0 for enterprise SSO

4. Configuration (Optional)

Edit ~/.medha/configs/config.json for advanced settings:

{
  "auth": {
    "type": "local"
  },
  "server": {
    "host": "localhost",
    "port": 8080
  },
  "database": {
    "type": "sqlite",
    "sqlite_path": "~/.medha/db/medha.db"
  },
  "git": {
    "sync_interval_minutes": 60
  },
  "security": {
    "token_ttl_hours": 24
  },
  "embeddings": {
    "enabled": false,
    "provider": "openai",
    "model": "text-embedding-3-small",
    "api_key_env": "OPENAI_API_KEY",
    "dimensions": 1536
  }
}

See Configuration Guide for all options including SAML setup.

5. Enable Semantic Search (Optional)

For AI-powered semantic search, enable embeddings:

# Via command line
./bin/medha --enable-embeddings

# Or set in config.json
"embeddings": { "enabled": true }

Requirements:

• Set OPENAI_API_KEY environment variable
• Or configure a compatible embedding API endpoint

Docker with embeddings:

{
  "mcpServers": {
    "medha": {
      "command": "docker",
      "args": [
        "run", "-i", "--rm",
        "-v", "/Users/yourname/.medha:/home/medha/.medha",
        "-e", "ENCRYPTION_KEY=your-32-char-key",
        "-e", "OPENAI_API_KEY=sk-your-key",
        "-e", "MEDHA_EMBEDDINGS_ENABLED=true",
        "tejzpr/medha-mcp"
      ]
    }
  }
}

MCP Tools

Medha uses human-aligned tools that express intent rather than implementation. This makes them easier for LLMs to use correctly.

medha_recall

"What do I know about X?" - Find and retrieve information:

{
  "topic": "authentication approach",
  "limit": 10
}

Also supports list_all: true for browsing and exact: "text" for literal search.

medha_remember

"Store this for later" - Create or update memories:

{
  "title": "Project Alpha Kickoff",
  "content": "# Meeting Notes\n\nDiscussed project timeline...",
  "tags": ["project", "meeting"],
  "replaces": "old-meeting-notes"
}

Use replaces to supersede old information (marks it as outdated).

medha_history

"When did I learn about X?" - Temporal queries:

{
  "slug": "project-alpha",
  "show_changes": true,
  "since": "7d"
}

medha_connect

"These are related" - Link or unlink memories:

{
  "from": "project-alpha",
  "to": "contact-john-doe",
  "relationship": "person"
}

medha_forget

"No longer relevant" - Archive a memory:

{
  "slug": "old-decision"
}

medha_restore

"Bring back that archived memory" - Undelete:

{
  "slug": "old-decision"
}

medha_sync

Manual sync to GitHub:

{
  "force": false
}

Memory Format

Memories are stored as Markdown files with YAML frontmatter:

---
id: project-alpha-kickoff-2024-01-15
title: "Project Alpha Kickoff Meeting"
tags: [project, meeting]
created: 2024-01-15T10:30:00Z
updated: 2024-01-15T14:00:00Z
associations:
  - target: contact-john-doe
    type: person
    strength: 1.0
---

# Project Alpha Kickoff Meeting

## Attendees
- John Doe (Project Lead)

## Key Decisions
...

Directory Structure

~/.medha/
├── configs/
│   └── config.json              # User configuration
├── db/
│   └── medha.db                 # System database (users, auth, repos)
└── store/
    └── medha-{username}/        # User's git repository
        ├── .medha/
        │   └── medha.db         # Per-user database (memories index)
        ├── 2024/
        │   └── 01/              # Date-organized memories
        ├── tags/
        │   └── meetings/        # Tag-based organization
        └── archive/             # Soft-deleted memories

Database Architecture (v2):

• System DB (~/.medha/db/medha.db): Users, authentication, repository registry
• Per-User DB (store/medha-{user}/.medha/medha.db): Memory index, associations, tags (git-tracked)

Development

Note: CGO is required for sqlite-vec. Ensure GCC is installed.

# Build
make build              # Build binary (CGO enabled)
make build-native       # Build optimized for current platform

# Run tests
make test               # Run unit tests
make test-integration   # Run integration tests
make test-functional    # Run functional tests
make test-all           # Run all tests

# Run server
make run                # stdio mode (MCP)
make run-http           # HTTP server mode
make run-with-embeddings # With semantic search

# Code quality
make lint               # Run linter
make vet                # Run go vet
make coverage           # View test coverage

# Docker
make docker-build       # Build Docker image
make docker-run         # Run container
make docker-run-with-embeddings # With embeddings

Database Rebuild

If you need to rebuild the database index from your git repository:

# Rebuild system database from git
medha --rebuilddb

# Rebuild and overwrite existing data
medha --rebuilddb --force

# Rebuild per-user database (v2)
medha --rebuild-userdb all           # All users
medha --rebuild-userdb username      # Specific user
medha --rebuild-userdb /path/to/repo # By path
medha --rebuild-userdb all --force   # Force overwrite

Contributing

Contributions are welcome! Please ensure:

• All tests pass (make test)
• Code is properly formatted (go fmt)
• Linter passes (make lint)
• Documentation is updated

License

This project is licensed under the Mozilla Public License 2.0.