Mem0
Mem0 is an open source memory layer for AI agents and LLM apps — it extracts the durable facts from each conversation, stores them, and pulls the relevant ones back into the prompt so your agent remembers users and context across sessions instead of starting cold every call.
Mem0 is an open source memory layer for AI agents and LLM apps — it extracts the durable facts from each conversation, stores them, and pulls the relevant ones back into the prompt so your agent remembers users and context across sessions instead of starting cold every call.
What is Mem0?
Mem0 is a memory layer that sits between your AI agent and a vector store, giving LLM applications long-term recall. After each interaction it uses an LLM to extract the facts worth keeping — a user’s preferences, decisions, prior context — stores them, and on the next call retrieves only the relevant memories to inject into the prompt. It ships as a Python and TypeScript SDK, a self-hostable server, or a managed cloud.
What is Mem0 best for?
Mem0 is best for developers building chatbots, support agents, or AI assistants that need to remember users across sessions without re-sending an entire conversation history every time. It’s a strong fit when you already have an agent and want to bolt on personalization and recall with a few lines of code, rather than designing your own storage, extraction, and retrieval pipeline from scratch.
What can Mem0 do?
- Extract and store memories automatically — one LLM call per interaction pulls out the durable facts and writes them to a vector store, so you don’t hand-manage what to remember.
- Multi-level memory — separate user, session, and agent-level memory so an assistant can track a person, a single conversation, and its own state independently.
- Hybrid retrieval — combines semantic (embedding) search, BM25 keyword matching, and entity linking to surface the most relevant memories.
- Graph memory — an optional graph store tracks relationships between entities for multi-hop reasoning (gated to the Pro cloud tier; available self-hosted).
- 19+ vector store backends — Qdrant is the default, with Postgres/pgvector, Chroma, Weaviate, Milvus, and many others supported.
- Model-agnostic — works with any LLM and embedding provider (OpenAI by default), so you’re not locked to one vendor.
- Framework integrations — native hooks for CrewAI, LangChain, Langflow, Flowise, and the Vercel AI SDK, among 20+ documented integrations.
Where does Mem0 fall short?
- Every memory write costs an LLM call. Extraction runs an LLM on each interaction, so at high volume the token and latency cost of writing memories is real — it’s not a free key-value store, and you’re paying your model provider per add.
- Graph memory is the paid hook on the cloud. On the managed platform, entity/relationship tracking is locked to the Pro tier ($249/mo), so teams that need graph memory jump straight past the cheaper plans — though self-hosting gives you graph memory without that gate.
- Temporal reasoning was historically weaker than specialist rivals. Earlier versions trailed tools like Zep and Hindsight on temporal-reasoning benchmarks; an April 2026 algorithm update closed much of the gap, but if precise time-ordered recall is your core need, benchmark it before committing.
Is Mem0 free?
Yes to self-host — the core library and server are Apache-2.0 and free to run on your own infrastructure with your own vector store and LLM keys (you still pay your model provider for the extraction calls). The managed Mem0 Platform is open-core: a free Hobby tier (10,000 add and 1,000 retrieval requests/month), then Starter at $19/mo, Pro at $249/mo (adds graph memory and consolidation), and custom Enterprise pricing for SSO, audit logs, and on-premise.
FAQ
Is Mem0 open source? Yes. The core memory layer is released under the Apache-2.0 license on GitHub, and you can self-host the full server. The hosted Mem0 Platform is a separate commercial product built on top of it.
Can I self-host Mem0 for free? Yes. Run it as a Python/Node library or as a Docker-based server backed by a vector store like Qdrant or Postgres/pgvector — no license fee. You still pay whichever LLM provider handles memory extraction and embeddings.
How is Mem0 different from a plain vector database? A vector store like Milvus or Weaviate stores and searches embeddings; Mem0 is a layer above it that decides what to remember, extracts facts with an LLM, links entities, and retrieves the right memories at query time. It uses a vector DB underneath rather than replacing one.
How does Mem0 compare to Letta? Both add long-term memory to agents. Letta is a full stateful-agent framework and server (the MemGPT successor); Mem0 is a lighter memory layer you drop into an agent you’ve already built. Pick Letta to run agents as persistent services, Mem0 to add recall to an existing stack.
What do I need to run Mem0? Python 3.9+ or Node.js, an API key for an LLM and embedding provider (OpenAI by default, or any supported model), and a vector store (Qdrant runs out of the box). The self-hosted server ships as a Docker image.