Mem0 — Memory Layer for AI Applications

The context-window problem, solved at the storage layer

A fresh LLM session knows nothing about you. Every conversation starts from zero — your name, your stack, your preferences, all re-introduced every time. Mem0 inserts a persistence layer between your application and the LLM so that conversational state, user facts, and long-term context survive sessions.

The 10-line integration

from mem0 import Memory
m = Memory()

# Remember something
m.add("Alice prefers TypeScript over JavaScript", user_id="alice")
m.add("Alice's project uses PostgreSQL and Redis", user_id="alice")

# Retrieve for the next LLM call
hits = m.search("what database does alice use?", user_id="alice")
# [{'memory': 'Uses PostgreSQL and Redis', 'score': 0.95}]

That's it. Mem0 handles extraction, deduplication, embedding, and semantic retrieval under the hood. Supports OpenAI, Claude, Gemini, and any LiteLLM-compatible model.

Architecture in 30 seconds

1. Fact extraction — an LLM extracts atomic facts from each interaction (e.g., "user prefers X").
2. Dedup & update — new facts are compared to existing memory; duplicates collapse, contradictions trigger updates.
3. Storage — facts stored as text + embedding in your vector DB (Qdrant, Pinecone, pgvector, Chroma).
4. Retrieval — at query time, the top-k relevant facts are pulled and injected into the LLM prompt.

Self-hosted vs cloud

Benchmarks from the Mem0 team show 26% accuracy improvement on LOCOMO long-conversation benchmark vs OpenAI's built-in memory, and 91% lower latency than re-feeding full history every call.

Real production use cases

• Personal AI assistants — Perplexity-style AI remembers your research topics across days.
• Customer support bots — agent recalls ticket history without SQL queries.
• Voice assistants — continuity across phone calls.
• Gaming NPCs — characters remember past player interactions.

Integration with popular frameworks

First-class support for LangChain, LangGraph, CrewAI, AutoGen, LiveKit, Vercel AI SDK, and the OpenAI Agents SDK.

Common pitfalls

1. Over-memorization — if you add every message, retrieval gets noisy. Use add selectively or rely on the auto-extraction heuristics.
2. Vector DB cold start — Qdrant/pgvector need index warm-up. First query after idle can take 500ms+.
3. Cost control — fact extraction uses an LLM call; budget accordingly. Claude Haiku or GPT-4o-mini are cost-effective for the extraction step.
4. User ID scoping — memories are keyed by user_id. Always pass it; default scoping leaks memories across users.