turbomem

· Documentation · Site

Embedded agent memory for TypeScript. Local-first by default — PGlite on disk or IndexedDB in the browser. Pluggable to edge and serverless with Upstash Vector or Pinecone. Runs in-process, no separate memory server, no Python sidecar.

Deployment

See the Edge guide for serverless setup.

Features

• Embedded runtime: in-process memory with no HTTP hop per call
• Semantic search: LLM fact extraction plus vector embeddings (OpenAI, local WASM, Voyage AI, or Google Gemini)
• PGlite storage: WASM Postgres with pgvector; data stays on disk in your app (default)
• sqlite-vec storage: optional SQLite backend via better-sqlite3 + sqlite-vec
• Upstash Vector storage: optional HTTP backend for edge runtimes (Cloudflare Workers, Vercel Edge)
• Pinecone storage: optional HTTP backend for edge runtimes (serverless vector index)
• Scoped by user, agent, or session: multi-tenant friendly out of the box
• Memory deduplication: merge, smart replace, or skip overlapping facts on write (enabled by default)
• Framework adapters: Mastra and Vercel AI SDK integrations ship as separate packages

Install

npm install turbomem

Set OPENAI_API_KEY for the default OpenAI embeddings and fact-extraction stack. PGlite is included, no extra database setup. For the optional sqlite-vec backend: npm install better-sqlite3 sqlite-vec. For edge: npm install @upstash/vector or npm install @pinecone-database/pinecone - see the Edge guide.

Prefer another provider? Embeddings support OpenAI, local (transformers), Voyage AI (VOYAGE_API_KEY), and Google Gemini (GEMINI_API_KEY); fact extraction supports OpenAI, Anthropic, and Google Gemini (plus any OpenAI-compatible endpoint via a custom baseURL). See the Providers reference.

Quick start

import { TurboMemory } from "turbomem";

const memory = new TurboMemory({
  embeddings: "openai",
  storage: "pglite",
  extraction: { provider: "openai", model: "gpt-4.1-mini" },
  openai: { apiKey: process.env.OPENAI_API_KEY },
});

await memory.init();

await memory.add(
  [{ role: "user", content: "I love hiking and I'm training for a half marathon this fall." }],
  { userId: "user_123" },
);

const results = await memory.search("What outdoor activities is the user into?", {
  userId: "user_123",
  limit: 5,
});

for (const { memory: m, score } of results) {
  console.log(`[${score.toFixed(3)}] ${m.content}`);
}

await memory.close();

When similar facts are added again, turbomem deduplicates by default - merging with an LLM, smart-replacing when the new fact is more specific, or skipping duplicates. Configure via deduplication or disable with { enabled: false }. See Configuration.

Adapters

npm install @turbomem/mastra turbomem
# or
npm install @turbomem/vercel-ai turbomem ai
# or
npm install @turbomem/okf

Links

Full guides, configuration reference, and runnable examples:

Site: https://turbomem.dev
Docs: https://docs.turbomem.dev

Requirements

Node.js 20+ or Bun. TypeScript recommended.

License

Apache License 2.0 see LICENSE for details.