Tool

Mem0

Memory layer for AI agents. Stores user/session facts, decays unused memory, and surfaces relevant context per turn. Drop-in for OpenAI-style chat loops.

URL: https://github.com/mem0ai/mem0
Taxonomy: tools.memory.agent_memory
License: Apache-2.0
Hosting: hybrid
Pricing: freemium
Status: active
Language: python
Stars: 24,000
First release: 2023
Self-host difficulty: moderate
Tags: agent-memory, long-term-memory, openai-compatible

What it is

Mem0 wraps your LLM calls with a memory store. Each turn, it extracts candidate facts (“user prefers dark mode”), stores them with embeddings, decays older entries, and fetches relevant memories on the next call.

When to use

Long-running conversational agents where the user expects continuity.
Multi-session products (assistants, personal CRMs, journaling apps).

When not to use

One-shot RAG over a static corpus — that’s not what memory is for.
Strict-privacy workloads where you don’t want a long-term store at all.

How Mem0 compares

AI-generated editorial comparisons against nearest peers (glm-4.6). Cached at build time; regenerate via node scripts/build-comparisons.mjs.

vs LightRAGAI · cached

The core tradeoff is between non-destructive retrieval from a static text corpus versus longitudinal fact tracking for a persistent user identity. LightRAG excels when you need to query a complex knowledge base—like a codebase or technical documentation—where understanding how entities relate is more important than simple keyword matching. It builds a graph to connect concepts across documents, which prevents naive vector search from losing critical context in cross-referenced materials. However, running this pipeline just to remember a user’s preference is massive overkill; the computational cost of graph extraction is wasted if your data doesn’t have internal relationships to map.

Mem0 wins clearly for personalized agents and chatbots. Instead of indexing a static document set, it actively manages a user’s profile by extracting facts ("Alice uses Vim"), storing them, and surfacing them at relevant times. If you are building a personal assistant or CRM, this behavioral continuity is essential. Conversely, Mem0 is the wrong tool for querying a company wiki; it doesn't ingest documents, and asking it to perform document retrieval will fail where LightRAG succeeds.

vs QdrantAI · cached

Mem0 is a specialized memory layer for agents, whereas Qdrant is a general-purpose vector database for retrieval. Mem0 handles the lifecycle of conversational facts—extracting "user prefers dark mode," decaying unused context, and surfacing it automatically in chat loops. It is strictly for maintaining state in long-running assistants or personal CRMs, and it fails for static knowledge retrieval. Do not use it for one-shot RAG over a company PDF corpus, as it lacks the infrastructure for large-scale document chunking.

Qdrant wins for raw RAG stacks and GraphRAG because it prioritizes high-performance search over conversational state. Its filtering engine allows you to mix dense vectors with sparse BM25 data, which is critical when you need to constrain search by metadata (like "filter by date"). However, Qdrant requires manual orchestration; you must write your own extractors and ingestion pipelines. If you need a drop-in memory that manages itself, use Mem0. If you need to query millions of vectorized documents with complex filters, use Qdrant.