turbomemEmbedded agent memory for TypeScript

Why embedded memory? ​

Many agent-memory setups rely on a separate service, a Python server, hosted platform, or self-managed vector DB. For TypeScript apps that means another process, extra infra, and an HTTP hop on every memory call. turbomem runs entirely in-process as a library you npm install.

Embedded vs server-based ​

If you need a cross-language managed platform, a dedicated memory service may fit better. If you're shipping a TypeScript app and want memory as a library, that's turbomem.

How it works ​

messages → Extractor (LLM) → facts → Embeddings → vectors → PGlite
query    → Embeddings → vector search (scoped) → ranked results

• add(messages, scope): extracts discrete facts from conversation, embeds them, and stores each fact with its vector and scope.
• addFacts(facts, scope): store explicit fact strings directly, skipping LLM extraction.
• search(query, scope): embeds the query and returns the closest matching facts, filtered by scope.
• getAll(scope): list every memory in a scope (newest first). The CLI exposes this as turbomem list.
• delete(id) / deleteAll(scope): remove one memory or wipe a scope.
• Scoping: every memory is tagged with optional userId, agentId, and sessionId for multi-tenant isolation.

See Architecture for the pipeline and API reference for the full method list.

When turbomem fits ​

• TypeScript agents (Node, Bun, browser, serverless) that need memory without a sidecar
• Products embedding memory into an app, not operating a memory platform
• Node, browser, edge, or serverless — PGlite locally, Upstash/Pinecone on Vercel and Workers
• Framework integrations via @turbomem/mastra or @turbomem/vercel-ai when you already use those stacks

Quick example ​

Local (default) ​

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,
});

console.log(results.map((r) => r.memory.content));

await memory.close();

Vercel / serverless ​

import { TurboMemory } from "turbomem";

const memory = new TurboMemory({
  storage: "upstash-vector",
  upstashVector: {
    url: process.env.UPSTASH_VECTOR_REST_URL,
    token: process.env.UPSTASH_VECTOR_REST_TOKEN,
  },
  embeddings: "openai",
  extraction: { provider: "openai", model: "gpt-4.1-mini" },
  openai: { apiKey: process.env.OPENAI_API_KEY },
});

await memory.init();
// same add / search API — see the Edge guide for provider recommendations on edge runtimes

See Edge for full setup. Pinecone is also supported — see Storage.

Deploy on Vercel / Next.js ​

Next.js on Vercel?
├── Edge runtime → upstash-vector or pinecone (see Edge guide)
├── Node serverless → upstash-vector or pinecone recommended; pglite only with persistent volume
└── Client-side → turbomem/browser with idb://

The vercel-ai-chatbot example switches between PGlite (local dev) and Upstash (production) via TURBOMEM_STORAGE.

Explore the docs ​

• Getting started - install, env setup, and first memory
• Configuration - embeddings, storage, extraction, and scoping
• Browser - IndexedDB persistence in client apps
• Architecture - pipeline, adapters, and error handling
• CLI - manage memories from your terminal
• MCP - give Claude Desktop a local long-term memory
• Examples - runnable projects in the repo
• API reference - types, methods, and error codes
• Contact - questions, feedback, and bug reports

Contact ​

Questions, feedback, or want to report an issue? See the Contact page or email arneesh@turbomem.dev.