Skip to main content

​
Quickstart

This guide gets you from zero to persistent AI memory in your MCP client (Cursor, Claude Desktop, or any MCP-compatible host) in under 5 minutes.

​
Prerequisites

  • Node.js 18 or newer
  • pnpm (npm install -g pnpm)
  • An MCP-compatible client (Cursor or Claude Desktop)

​
1. Clone and build

git clone https://github.com/4stax-hq/kontxt.git
cd kontxt
pnpm install
pnpm build

​
2. Initialize your vault

node packages/cli/dist/index.js init
This creates a .kontxt/vault.db file — your local memory store. Nothing leaves your machine.

​
3. Configure your MCP client

Open your Cursor MCP settings and add:
{
  "mcpServers": {
    "kontxt": {
      "command": "node",
      "args": ["/path/to/kontxt/packages/mcp-server/dist/index.js"]
    }
  }
}
Replace /path/to/kontxt with your actual clone path.

​
4. Test it

Try the CLI to verify everything is working: 
# Store a memory
node packages/cli/dist/index.js add "I prefer TypeScript over JavaScript" --type preference

# Search your vault
node packages/cli/dist/index.js search "language preferences"
You should see your stored memory returned in the search results.

​
5. Use it in your AI client

Restart your MCP client. From now on, kontxt will automatically surface relevant memories from your vault into your AI sessions. Try asking your AI client something related to what you stored — it should already know.
You now have persistent memory running locally. Every session builds on the last.

​
Optional: embeddings

For better semantic search, set your OpenAI API key before building: 
export OPENAI_API_KEY=sk-...
pnpm build
Without it, kontxt falls back to keyword matching which still works for most use cases.

​
Next steps

How memory works

Understand relevance ranking and memory types

MCP tools reference

Full reference for all exposed MCP tools

CLI reference

All CLI commands and flags

Self-hosting

Run kontxt on your own server