Mem0

Long-term memory for LLM apps is the problem Mem0 targets: after each exchange it asks an LLM to pull out facts worth keeping, embeds them, and files them under a user, session or agent, so later prompts retrieve only the relevant history instead of replaying whole transcripts. The current algorithm appends new memories rather than rewriting old ones, links named entities across records, and at query time combines vector similarity, BM25 keyword scoring and entity matches. You get a Python package (mem0ai), a TypeScript SDK, and a Docker Compose server with a dashboard and API keys for team use; the separate CLI is documented against the hosted API. It needs an LLM and an embedding model, defaulting to OpenAI but working with Ollama, Anthropic, Groq and others, plus a vector store such as Qdrant, Chroma, pgvector, Milvus or FAISS, so a fully local setup is possible. The code is Apache 2.0; the company also sells a hosted platform, and its published benchmark scores include proprietary optimizations the open version lacks. Agent builders who want recall across sessions without designing a memory schema are the audience.