mem0: the dominant open-source memory layer for production AI agents (~40k stars)

Short version: mem0 is the dominant open-source memory layer for AI agents — ~40k GitHub stars, Python + TypeScript SDKs, first-class user_id scoping on every call, swappable vector backend (Pinecone, Qdrant, Weaviate). If you're building a product where every end-user gets their own memory, or you need production-grade retrieval, mem0 is the default pick. Apache-2.0 licensed. ~20-minute install. By mem0ai.

What is mem0?

mem0 treats memory as a database of structured facts about users. Every add() call extracts concepts with an LLM and stores them scoped to a user_id. Every search() call retrieves the most relevant memories for that user. Simple mental model. Production-grade infrastructure under the hood.

It's the most widely-adopted memory layer in open source — ~40k stars, deep SDK coverage in Python and TypeScript, broad community support, a managed cloud option for teams that don't want to run their own infrastructure.

Who this is for

Teams building AI products where every end-user needs their own agent with their own memory (coaching apps, CX agents, personal assistants, health apps).
Engineering teams that want an explicit API — your code controls when memory is written and retrieved.
Companies that need per-tenant isolation for enterprise customers.
Anyone building at production scale who wants the community-default pick.

Skip this if

You're a solo founder just experimenting — memory-mcp is a 2-minute install that proves the value before you invest in mem0's setup. Or if you want auto-ingest from your personal SaaS apps, look at supermemory.

What problem it solves

Most memory tools are designed for personal use — one user, one agent, one memory. That breaks the moment you build a product. Every end-user of your app needs their own scope. Cross-user leakage is a bug (worse, a compliance violation). Multi-tenancy is mandatory.

mem0 is built for that from day one. user_id is on every call. Memory is isolated by default. You can scale horizontally by creating more users, not more servers. The mental model — "structured user memory, explicit writes, semantic retrieval" — matches how a real product works.

How to install it (plain English)

Install the SDK. Python: pip install mem0ai. TypeScript: npm install mem0ai.
Set your LLM API key. OpenAI, Anthropic, or compatible.

Minimal script:

from mem0 import Memory

m = Memory()

m.add("Amara is the CX lead at Linea.", user_id="vedant")
results = m.search(query="who is on my team?", user_id="vedant")

Configure your vector store. Default is local; swap to Pinecone, Qdrant, or Weaviate for scale.
(Optional) Use as an MCP server. mem0 ships an MCP bridge for direct Claude Code / Cursor integration.

Full walkthrough: /memory/tools/mem0.

What you can do with it (for a non-technical founder)

If your team is building a product with AI memory:

Every user gets their own brain — scoped by user_id, no cross-user leakage.
Production-ready retrieval — semantic search with graph augmentation, not just string match.
Swap backends as you scale — start local, move to Pinecone/Qdrant/Weaviate when you need to.
Managed or self-hosted — mem0.ai offers a managed tier; you can also self-host the full stack on your VPC.
Largest ecosystem — 40k+ stars means the most docs, most examples, most Stack Overflow coverage.

What CLO adds on top

mem0 gives your users persistent memory inside your product. Cognition CLO gives your internal team a retention layer on top of your organizational knowledge. Different scopes: mem0 is for your users' memory; CLO is for your team's retention.

Some teams run both: mem0 for the consumer memory inside their product, CLO for the internal knowledge retention of their employees.

FAQ

Does mem0 require an LLM API to work?

Yes — the extraction step uses an LLM to decide what's worth storing. Cheap models (GPT-4o-mini, Haiku) work fine for extraction.

Can I control what gets extracted?

Yes. The extraction prompt is configurable. For stricter control, you can bypass the LLM step and call add() with pre-structured data.

How does multi-user scoping work?

Every add() and search() call takes a user_id. Memories are stored keyed by that ID. Queries with the wrong ID return nothing.

Can I run mem0 without a dedicated vector database?

Yes — the default setup uses a local store that works fine for under ~1000 users. Swap to Pinecone / Qdrant / Weaviate when you scale.

Is the cloud version HIPAA compliant?

The self-hosted version can be deployed on your own HIPAA-compliant infrastructure. The managed cloud is not inherently HIPAA-covered — check mem0.ai's current enterprise offerings.

How does it compare to Letta?

mem0 is a memory library scaled across users. Letta is an agent framework with deep context management for single agents. For multi-user products, mem0. For single-agent deep context, Letta. See /blog/letta-vs-mem0 for the full comparison.

How does it compare to memory-mcp?

memory-mcp is the simplest-thing-that-works (solo use). mem0 is production-grade (multi-user). See /blog/mem0-vs-memory-mcp.

Ready to install? Full walkthrough at /memory/tools/mem0. Comparisons: vs memory-mcp · vs Letta · vs supermemory. Credit to @mem0ai and the community maintainers — star the repo if mem0 ends up powering your production memory.