recall-kit

Recall Kit is an MCP server that enables AI agents to query, retrieve, and contribute experiential knowledge for AI development. It acts as a centralized experience repository, allowing the agent to fetch relevant past solutions to similar problems and to submit new experiences after solving issues. The MCP integration lets your AI agent transparently search across stored experiences using natural language prompts and vector/text search capabilities, reducing repeated investigation and speeding up development. The included web UI provides a browsable interface for managing and exploring experiences, while the MCP server component handles protocol-level queries and submissions from AI agents. To use it, point your agent at the Recall Kit MCP endpoint and pass prompts that reference related issues or tasks; the agent can then retrieve matching experiences or push new learnings back into Recall Kit when it solves problems.

Prerequisites:

• Node.js 18+ (for the MCP server)
• npm (or yarn)
• PostgreSQL database (for Recall Kit data storage)
• OpenAI API Key (optional, required for vector search if you enable embedding features)

1. Clone the repository and navigate to the project root:

git clone <repository-url>
cd recall-kit

2. Install dependencies for the MCP server (and other subprojects if needed):

cd mcp-server
npm install

3. Configure environment variables for the MCP server:

• Create a .env file in mcp-server with necessary configurations (e.g., database connection, OpenAI keys if used, and any server-specific settings). See mcp-server/README.md for exact variables.

4. Prepare the PostgreSQL database:

# Example (customize per your setup)
createdb recall_kit
psql -d recall_kit -f database_init.sql

5. Run the MCP server (and any other required services):

# Start MCP Server
cd mcp-server
npm run dev

6. Confirm the MCP endpoint is reachable:

• MCP Server: http://localhost:3001/mcp (or the port configured in your setup)
• If using the public URL, ensure network access and API key validity.

Note: If you also want to run the web UI locally, follow the web subproject README for its own setup and run commands.

Tips and considerations:

• The Recall Kit MCP server supports both querying and submitting experiences. When your agent cannot find an existing experience, you can choose to submit a new one for future reuse.
• If you enable vector embeddings for search, an OpenAI API key is typically needed; ensure your key has access to the embedding models you plan to use.
• The API key parameter in the MCP endpoint can be optional for read-only queries. If omitted, the server may still allow retrieval but will not permit submissions of new experiences.
• The project includes a web UI for browsing and managing experiences; you can reference the web/README.md for UI-specific setup details.
• Ensure proper database migrations are applied in order if you’re setting this up from scratch; the supplied migrations seed schemas and indices necessary for efficient search.
• When upgrading components, check MCP protocol changes or version notes in the underlying mcp-server README to maintain compatibility with your AI agent integrations.