Skip to content

Quickstart

You’ll need a Memoturn API key and project slug. Get them from memoturn.ai → API keys in your project.

1. Pick a transport

pnpm add @memoturn/sdk

import { Memoturn } from "@memoturn/sdk";

const mt = new Memoturn({
  apiKey: process.env.MEMOTURN_API_KEY!,
  projectId: "my-project",
});

await mt.recordTurn({ role: "user", content: "deploying the worker tonight" });

const { hits } = await mt.searchMemory({ query: "deploy", k: 5 });
console.log(hits);

pnpm add -g @memoturn/cli
mt init                                    # one-time: store key + project
mt record "deploying the worker tonight"
mt search "deploy"

{
  "mcpServers": {
    "memoturn": {
      "transport": "http",
      "url": "https://api.memoturn.ai/v1/mcp",
      "headers": {
        "Authorization": "Bearer ${MEMOTURN_API_KEY}",
        "X-Project-Id": "my-project"
      }
    }
  }
}

Restart Claude Code; the record_turn, search_memory, etc. tools will appear under the memoturn server.

2. Search modes

search_memory accepts a mode:

mode	use it for
`auto` (default)	Hybrid recall — dense + lexical + DO-hot fused via RRF.
`chunks`	Same hybrid pipeline, but excludes rolling summaries.
`summaries`	Only retrieve consolidated session summaries. “What did we figure out?”
`entities`	Symbol/file/error-code lookup. Pair with the regular search for full text.
`code`	Hybrid plus a fourth fusion leg that boosts turns matching extracted symbols.

await mt.searchMemory({ query: "auth flow", mode: "summaries" });

for await (const event of mt.subscribe()) {
  console.log(event.kind, event.payload);
  // turn_recorded · memory_pinned · memory_forgotten · broadcast_set · presence_updated · consolidation_completed
}

Next

MCP tool reference — every tool, every input.
SDK guide — typed methods + WebSocket subscribe.
Architecture — what’s between you and Postgres.