Long-term memory for your AI agents
Memory, knowledge, learnings, and skills for AI agents. Persistent context that survives across sessions, providers, and platforms — open-source and yours to run. Connect over a REST API, the TypeScript SDK, the CLI, or the built-in MCP server.
The hosted SwarmRecall service has been discontinued. The full stack is open-source: clone it, run it, own your data.
Run your own instance
Self-hosting takes three steps. Everything runs locally with Docker — no managed service required. The self-hosting guide covers credentials and deployment.
Clone the repo
Pull SwarmRecall from GitHub. It is a pnpm + Turborepo monorepo: API, dashboard, SDK, CLI, and MCP server.
Run the stack
docker compose up brings up Postgres + pgvector and Meilisearch. pnpm dev starts the API on :3300.
Point your agents at it
Set SWARMRECALL_API_URL to your instance. The SDK, CLI, and MCP server default to http://localhost:3300.
Four pillars of agent intelligence
Everything your agents need to build long-term understanding.
// Memory
Semantic vector search across everything your agent has ever stored. Full-text and embedding-based retrieval with automatic deduplication.
// Knowledge
Build a knowledge graph that connects entities, concepts, and relationships. Your agents understand context, not just keywords.
// Learnings
Pattern extraction from agent interactions. Auto-distill repeated successes and failures into reusable insights.
// Skills
A registry of agent capabilities. Track what your agents can do, version their skills, and share across your swarm.
// Shared pools
Let several agents read and write the same memories, entities, learnings, and skills, with read or read-write access per module.
// Dreaming
Consolidation cycles deduplicate memories, decay stale ones, merge duplicate entities, and summarize finished sessions.
// Dashboard
A web dashboard on your own instance for browsing memories, the knowledge graph, learnings, skills, and API keys.
Memory MCP server
Give any MCP-compatible agent direct access to its memory. The server exposes 53 tools and 4 resources across memory, knowledge, learnings, skills, shared pools, and dream cycles.
- Local stdio
- Run
swarmrecall mcpfrom the CLI. Desktop and editor clients start it as a subprocess. - Remote HTTP
- Point the client at your instance's
/mcpendpoint over Streamable HTTP and authenticate with your agent's API key.
{
"mcpServers": {
"swarmrecall": {
"command": "swarmrecall",
"args": ["mcp"],
"env": {
"SWARMRECALL_API_KEY": "sr_live_...",
"SWARMRECALL_API_URL": "http://localhost:3300"
}
}
}
}Simple SDK, powerful results
A few lines of code give your agent permanent memory — pointed at your own instance. See the SDK guide for every namespace.
// Initialize SwarmRecall
import { SwarmRecallClient } from '@swarmrecall/sdk';
const recall = new SwarmRecallClient({
apiKey: process.env.SWARMRECALL_API_KEY,
baseUrl: process.env.SWARMRECALL_API_URL, // → http://localhost:3300
});
// Store a memory
await recall.memory.store({
content: "User prefers dark mode",
category: "preference",
});
// Search memories semantically
const results = await recall.memory.search("user preferences");// open source
MIT licensed. Self-host it, fork it, make it yours.