Mem0

Why it's ranked here

Apache 2.0, and genuine architectural range: one project ships a plain library, a self-hosted server, a hosted platform, and integrations across a long list of vector databases, not a thin wrapper around one backend. The code defends the multi-tenant model it advertises, refusing to let free-form metadata silently move a memory into someone else's scope. Where caution belongs is candor: the vendor's own documentation says its best benchmark numbers come from the managed platform's proprietary layer, not the code anyone can run themselves. That is honest, and it means judging the open code and the paid tier as two separate products.

What's good

The scope model is enforced in code, not just documented: identifiers a caller passes as free-form metadata get stripped if they collide with the fields that actually control ownership, so a memory can't be reassigned to another user by accident. Secrets embedded in configuration objects are redacted before anything reaches the bundled analytics pipeline. Vector store support is unusually broad, covering more than a dozen backends as optional installs, so a team already running Postgres, Redis, or Elasticsearch can point memory storage at infrastructure it already operates.

Tradeoffs

Both bundled defaults point at one vendor: fact extraction and embedding generation use OpenAI models unless reconfigured, and the self-hosted server's list of ready providers covers only three vendors without a rebuild. That server also demands deliberate auth setup, teams upgrading from an older, authless build get locked out until an admin key is set or auth is explicitly disabled. And the project's own quickstart is written for an AI agent to create its account and mint an API key with no human present, a design choice worth noticing before wiring this into anything touching real user data.

How to use it well

Start with the plain library for prototyping, move to the self-hosted server once a team needs its own dashboard, auth, and infrastructure, and use the managed platform for zero ops and the strongest retrieval quality, since that quality is not fully available in the open code. For real hybrid-search gains, install the optional language-processing extras and swap in a stronger embedding model, the bundled default is a small one. It suits products that need to remember a specific person across sessions; it is a poor fit for teams that need a fully local, single-vendor-free stack out of the box.

Observed

License

License: Apache 2.0 (pyproject.toml and README.md)

Language

Primary language Python, requires Python >=3.10,<4.0 (pyproject.toml)

Packaging

Packaged as the PyPI package mem0ai and the npm package mem0ai; also distributable as a self-hosted Docker Compose stack (server/main.py) and a hosted cloud platform

Interfaces

Interfaces: a Python library (Memory / AsyncMemory / MemoryClient / AsyncMemoryClient exported from mem0/__init__.py), a CLI, and a self-hosted FastAPI REST server (server/main.py)

Authentication

Self-hosted server refuses to start without JWT_SECRET unless AUTH_DISABLED=true is set (server/main.py)

Vector store integrations

Optional vector-store integrations ship as extras: qdrant, chroma, pinecone, weaviate, elasticsearch, milvus, redis/redisvl, pgvector-compatible Postgres, MongoDB, Oracle, Azure AI Search, and others (pyproject.toml)

Default providers

Default LLM is OpenAI gpt-5-mini and default embedder is OpenAI text-embedding-3-small (README.md, server/main.py)

Bundled provider restrictions

Bundled self-hosted server restricts usable providers to openai, anthropic, and gemini for LLMs and openai, gemini for embedders without rebuilding the image (server/main.py)