MCP Vector Memory

Persistent vector memory for AI coding agents. Give your AI agents long-term memory that survives across conversations.

Works with Antigravity · VS Code / Copilot · Claude Code · Codex

The Problem

AI coding agents forget everything between conversations. Every new session starts from zero — they repeat questions, lose architectural decisions, and forget how your project is set up.

The Solution

MCP Vector Memory gives your agents persistent, semantic memory. Agents can save and search memories using natural language. Memories are stored locally in SQLite with vector similarity search — no API keys, no cloud, no cost.

Agent: "I need to set up the database"
  → search_memory("database setup") 
  → Returns: "PostgreSQL 18 with pgvector, HNSW index, user=app_user..."
  → Agent works WITH full context from past sessions

Quick Start

1. Install

pip install mcp-vector-memory

2. Configure Your IDE

Add to ~/.gemini/antigravity/mcp_config.json:

{
  "mcpServers": {
    "memory": {
      "command": "mcp-vector-memory",
      "args": [],
      "env": {}
    }
  }
}

Create .vscode/mcp.json in your workspace:

{
  "mcp": {
    "servers": {
      "memory": {
        "command": "mcp-vector-memory",
        "args": [],
        "env": {}
      }
    }
  }
}

Add to ~/.claude/claude_desktop_config.json:

{
  "mcpServers": {
    "memory": {
      "command": "mcp-vector-memory",
      "args": [],
      "env": {}
    }
  }
}

Add to your Codex MCP configuration:

{
  "mcpServers": {
    "memory": {
      "command": "mcp-vector-memory",
      "args": [],
      "env": {}
    }
  }
}

3. Restart your IDE

That's it. Your agents now have persistent memory.

Tools Available

How It Works

┌─────────────────┐     ┌──────────────────────┐     ┌─────────────┐
│   AI Agent      │────▶│  MCP Vector Memory   │────▶│   SQLite    │
│  (any IDE)      │◀────│  (local process)     │◀────│  + vectors  │
└─────────────────┘     └──────────────────────┘     └─────────────┘
                              │
                         Embeddings
                      (all-MiniLM-L6-v2)
                        runs locally

1. Agent calls save_memory → text is embedded locally → stored in SQLite with vector index

2. Agent calls search_memory → query is embedded → SQLite finds most similar memories

3. Everything runs locally. No API calls, no cloud, no cost.

Configuration

All configuration is via environment variables:

PostgreSQL Backend (Optional)

For production deployments with PostgreSQL + pgvector:

pip install mcp-vector-memory[postgres]

MCP_MEMORY_BACKEND=postgres
PGHOST=localhost
PGPORT=5432
PGUSER=mcp_memory
PGPASSWORD=your_password
PGDATABASE=mcp_memory

Docker (PostgreSQL)

docker compose up -d

See docker-compose.yml for the full setup.

OpenAI Embeddings (Optional)

For higher quality embeddings via OpenAI API:

pip install mcp-vector-memory[openai]
export OPENAI_API_KEY=sk-...
export MCP_MEMORY_EMBEDDING_PROVIDER=openai

Contributing

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

License

MIT — see LICENSE