Exocortex provides a local-first memory store with embeddings, full-text indexing, and a unified retrieval API exposed through an MCP server. It supports a wide range of tools via the MCP protocol, including memory_store, memory_search, memory_get, memory_update, memory_context, memory_browse, memory_feedback, memory_entities, memory_timeline, memory_ingest, memory_link, and more. The server runs locally and communicates over stdio so you can connect a variety of agents and CLIs that implement the MCP interface. Use it to persist memories across sessions, enrich them with embeddings, and retrieve them with hybrid ranking (RRF + BM25) and graph-aware expansion for richer context. The system also exposes a REST API and a React dashboard for browser-based management, plus a CLI for direct operations.

Prerequisites:

• Node.js >= 20
• pnpm (package manager)
• Git

Installation steps:

1. Clone the repository: git clone https://github.com/shawnhack/exocortex.git
2. Navigate to the project root: cd exocortex
3. Install dependencies and build (first build downloads the embedding model): pnpm install pnpm build
4. Start the MCP server via the MCP client/agent integration (see Quick Start) or connect via an MCP client using the path to the built MCP entrypoint, e.g.:

   Example for an external MCP client when configured to invoke the server via node

   node /path/to/exocortex/packages/mcp/dist/index.js
5. Open the dashboard (if using the REST API and dashboard): http://localhost:3210

Notes:

• The exact command to launch the MCP server may vary depending on your integration (CLI, VS Code, Claude, Gemini, etc.). The recommended integration uses the node entrypoint exposed by the @exocortex/mcp package.

Tips and considerations:

• The server runs embeddings locally using HuggingFace transformers—no external API keys required for embeddings.
• It stores memories in SQLite with a vector store for embeddings and a BM25 index for full-text search.
• The MCP tools support a wide range of operations including ingestion (markdowns), graph operations (memory_link, memory_graph), maintenance tasks (memory_maintenance), and knowledge management (memory_consolidate, memory_diff).
• Environment variables are typically optional; you can override defaults via your MCP client config or the server’s environment. Typical concerns include memory usage for embeddings and model sizes—first build downloads the embedding model (~80MB).
• If you encounter issues with connections from MCP clients, ensure the server process is reachable and that the path to the MCP entrypoint is correct in your integration configuration.