Fully embedded
Runs inside your Node, Bun, or browser process. No sidecar server, no HTTP hop per memory call.
Agent memory for TypeScript
Local-first memory in your process
Extract facts, embed them, and recall scoped memories from your app process, with pluggable vector storage and framework adapters. npm install turbomem. No Python sidecar. No memory server.
INP
Input
addFacts()
EXT
v2Extractor
LLM
EMB
Embed API
768-d vector
DED
v2Dedup
cosine ≥ 0.92
STO
Storage
vector backend
SEA
Search
cosine KNN
RES
Result
scoped rank
Input
"User loves trail running in the mountains"
TypeScript
Native library
In-process
No HTTP hop
Apache-2.0
Open source
Semantic
LLM facts + vector search
Open source today. Cloud soon.
Self-host the full library now, or join the waitlist for managed turbomem Cloud.
Open source
Free · self-hosted
Apache-2.0 · full library access
- PGlite, sqlite-vec, and Upstash storage adapters
- LLM extraction, deduplication, and scoped search
- Browser, Node, Bun, and edge runtimes
turbomem Cloud
Coming soon
Managed API, dashboard, and team features
- Hosted memory API with SLAs
- Usage analytics and observability
- Enterprise-ready controls
Built for embedded agent memory
Everything you need for durable agent memory, semantic recall, tenant scoping, and storage that fits your runtime.
Semantic search
LLM fact extraction and embeddings, with cosine-ranked recall. Swap embedding providers without changing your app code.
Pluggable storage
PGlite + pgvector by default. Optional sqlite-vec for SQLite-native apps, or Upstash Vector for edge runtimes.
Adapter-based
Swap embeddings, storage, and framework adapters like Mastra, Vercel AI SDK, and more.
Type-safe
Strict TypeScript with Zod-validated inputs. Node 20+, Bun, browser, serverless.
Terminal CLI
Add, search, and manage memories from your shell with one-shot commands or a REPL.
Browser-ready
IndexedDB-backed PGlite in the browser, durable client-side memory with no remote database.
Edge-ready
Upstash Vector over HTTP on Cloudflare Workers and Vercel Edge. Same API, different backend.
Multi-tenant scoping
Tag memories with userId, agentId, sessionId. Search and list respect scope out of the box.
How it works
Conversations become facts. Facts become vectors. Queries return ranked memories, all in-process.
01
Extract
An LLM turns chat into discrete, third-person facts about the user.
02
Embed & store
Embeddings land in PGlite, sqlite-vec, or Upstash pick local Postgres, SQLite, or HTTP for the edge.
03
Retrieve
Semantic search returns scoped results with cosine similarity scores.
turbomem Cloud
Join the waitlist for managed agent memory
Hosted API, dashboard, and team features - without giving up the open-source library you can self-host today.
Join waitlist