mem0

Type: full-code · Vendor: Mem0 · Language: Python, TypeScript · License: Apache-2.0 · Status: active · Status in practice: mature

Links: homepage docs repo

Drop-in memory layer that extracts salient facts from agent conversations and serves them back as personalized context across sessions, users, and agents.

Description. Mem0 is a managed and open-source memory layer for AI agents that turns ongoing conversations into a stored set of user facts, preferences, and episodes which the agent can search on subsequent turns. The agent sends raw messages to `add`; an LLM extracts salient items, deduplicates them against prior memory, and writes them to a vector store scoped by `user_id`, `agent_id`, or `run_id`. On later turns the agent calls `search` to retrieve the most relevant memories instead of replaying the full chat history. The project ships Python and TypeScript SDKs plus a hosted platform, and is described in the paper 'Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory'.

Agent loop shape. Sidecar memory loop wrapped around an external agent. On every user turn the agent calls `mem0.search(query, user_id)` to retrieve relevant prior memories and injects them into the LLM prompt; after the turn the agent calls `mem0.add(messages, user_id)` which sends the messages through an extraction LLM that derives facts, resolves conflicts against existing memories, and writes the survivors to managed vector storage. Memory operations are add / search / update / delete; identity is keyed by user_id, agent_id, or run_id.

Primary use cases

persistent user memory across sessions
personalized assistants that learn preferences over time
reducing prompt bloat and token cost on long conversations
cross-agent shared memory in multi-agent systems

flowchart TD user[User turn] --> search[mem0.search user_id] search --> store[(Vector store)] store --> ctx[Relevant memories] ctx --> llm[Agent LLM] llm --> reply[Reply to user] reply --> add[mem0.add messages user_id] add --> extract[Extraction LLM] extract --> dedupe[Dedupe vs existing] dedupe --> store

Key concepts

add / search / update / delete (docs) — Four operations on the memory store; `add` extracts facts via LLM, `search` returns relevant memories.
user_id / agent_id / run_id → session-isolation (docs) — Scoping identifiers that key memories to a person, an agent, or a single run.
Memory extraction (infer=True) (docs) — An LLM pulls out key facts, decisions, or preferences from raw messages and stores those rather than the verbatim turns.
Short-term vs long-term memory → episodic-summaries (docs) — Short-term covers conversation history, working memory, attention; long-term covers factual, episodic, and semantic memory.
Vector store backend → vector-memory (docs) — Managed vector storage holds extracted memories; an entity-linking pass replaces the older optional graph store.