🧠 MCP Memory Pro: The Big Five

Production-Ready Knowledge Graph Infrastructure with Zero-Config Context Awareness.

@steevenz/mcp-memory-pro is a high-performance MCP server designed to transform your IDE into an enterprise-grade second brain. By consolidating complex graph operations into "The Big Five" tool suite and introducing automatic project context caching, it offers a seamless, sticky memory experience for AI agents.

🚀 Vision: Zero-Config Context

Unlike standard memory servers that require manual project ID management, @steevenz/mcp-memory-pro understands the filesystem.

1. Init Once: When you initialize a project, it writes a .cache/memory-pro/project.json file in your target project root.
2. Sticky Memory: Every subsequent tool call automatically resolves the project context by walking up the directory tree to find that cache file.
3. Unified Tools: It collapses ~20 legacy tools into 5 powerful primitives, saving your LLM's "cognitive budget" (context window) while increasing precision.

🛠️ The Big Five Tool Suite

1. init_project (The Anchor)

Purpose: Sets up the bridge between your filesystem and the memory graph.

• Action: Creates/Updates project ID and writes the local cache file.
• Automation: Automatically appends the cache folder to your .gitignore.

2. upsert_graph (The Writer)

Purpose: Efficiently record nodes and edges with historical tracking.

• Capabilities:
  • Batch insert/update nodes and relationships.
  • Atomic Observations: Append factual snippets without overwriting main content.
  • Version Tracking: Automatic incrementing of entity versions on every update.

3. query_graph (The Seeker)

Purpose: Extract knowledge using multiple traversal and search modes.

• Modes:
  • search: Full-text search (FTS5) with metadata filtering.
  • traverse: Neighborhood traversal (N-hops).
  • path: Shortest pathfinding between two entities.
  • list: Paginated retrieval of raw nodes (supports include_observations).
  • subgraph: Context extraction for LLM injection.

4. manage_data (The Maintainer)

Purpose: Keep your brain healthy and portable.

• Actions:
  • summarize: Request graph statistics.
  • maintenance: Optimize DB performance (VACUUM/Optimize).
  • purge: Reclaim physical space from deleted items.
  • backup: Create database backups.
  • restore: Restore from backups.
  • export: Dump the entire project graph as a JSON blob.

5. delete_data (The Cleaner)

Purpose: Clean up nodes, edges, or entire project graphs.

6. health_check (The Monitor)

Purpose: Check server and database health with integrity checks.

⚙️ Configuration & Setup

Quick Start with npx (Recommended)

Add the following to your MCP configuration (Claude Desktop, Cursor, Windsurf, etc.):

{
  "mcpServers": {
    "memory": {
      "command": "npx",
      "args": ["-y", "@steevenz/mcp-memory-pro"],
      "env": {
        "MCP_DATA_DIR": "./data",
        "MCP_LOG_LEVEL": "info"
      }
    }
  }
}

Local Development (Before Publishing)

For testing before publishing to npm:

git clone https://github.com/steevenz/mcp-memory-pro.git
cd mcp-memory-pro
npm install

Then use the local bin script:

{
  "mcpServers": {
    "memory": {
      "command": "node",
      "args": ["path/to/mcp-memory-pro/bin/memory-pro.js"],
      "env": {
        "MCP_DATA_DIR": "./data",
        "MCP_LOG_LEVEL": "info"
      }
    }
  }
}

Environment Variables (Optional)

• MCP_DATA_DIR: Override the default data storage (Default: ./data inside server root)
• MCP_LOG_LEVEL: Logging verbosity (debug, info, warn, error)

📂 Architecture

• SQLite Backend: Uses sql.js (WASM-based) for cross-platform compatibility with in-memory operations and debounced persistence.
• Safety Protocol: BFS traversals are capped at 2,000 nodes and 30 seconds to prevent resource exhaustion.
• Performance: Debounced write persistence (1-second batching) for optimal I/O.
• Telemetry: All tool calls log performance metrics (Execution time > 500ms triggers a warning).
• Health Monitoring: Built-in health check tool for database integrity verification.

Known Limitations

• FTS5 (full-text search) is not supported by sql.js and is skipped gracefully
• No concurrent writes (sql.js limitation)
• Database operations are in-memory with manual persistence
• No native connection pooling

📚 Documentation

• Setup Guide: See docs/guidelines/how-to-setup.md for universal setup instructions (Claude, Cursor, Windsurf, Trae, etc.)
• API Reference: See docs/api.md for detailed tool documentation

🤝 Contribution

Built for scalability, clarity, and production readiness.

[!TIP] Pro Tip: To switch contexts, simply move your terminal to a different project and run init_project. The memory follows your directory focus.