What it is
Mem0 gives an agent long-term memory. It pulls facts and preferences out of conversations, stores them, and retrieves the relevant ones in later sessions so the agent does not start from zero.
The open-source SDK for Python and TypeScript runs as a library or as a self-hosted server with Postgres and pgvector, and calls an LLM (OpenAI gpt-5-mini by default) on every add. The April 2026 release of the SDK switched to add-only extraction and removed graph memory from open source; graph memory is now a feature of the paid Mem0 Platform.
Mem0 publishes results on LoCoMo (92.5), LongMemEval (94.4) and BEAM with an open evaluation harness, but says the headline numbers are for its managed platform, which includes optimisations the SDK lacks.
Our review
Mem0 is the default memory layer for agents, with about 67,700 GitHub stars, more than any rival, and an API that amounts to add and search calls around an existing loop. It also publishes an open evaluation harness for its benchmark claims.
The April 2026 algorithm changed how it behaves. Extraction is now one LLM call that only adds memories and never updates or deletes them; retrieval fuses semantic, BM25 keyword and entity matching with time-aware ranking. Mem0 reports LoCoMo rising from 71.4 to 92.5, but those are platform figures, and the README says open-source users should expect similar direction rather than identical numbers. Add-only storage also means old facts are not overwritten, so a user's former job sits beside the new one and relies on ranking to pick correctly.
The open-core split has sharpened. Graph memory was removed from the SDK in the April 2026 release and is now Platform-only, starting with the $249-a-month Pro plan. On the hosted side the binding limit is retrievals rather than storage: Starter at $19 includes 50,000 adds but only 5,000 searches a month, and an agent that searches on every turn uses that in days.
The SDK sends usage telemetry to PostHog by default, sampling hot-path events at 10%, until MEM0_TELEMETRY is set to false, and a cross-tenant conversation leak in session scoping was disclosed on 20 October 2026. Open issues ask for tagged Docker images and note that keyword search and entity extraction assume English, which matters for any non-English support desk.
Where it earns its place
A product engineer adding personalisation to a consumer chat app with 20,000 monthly users
Remember each user's preferences between sessions
Call add after each conversation and search before each reply. On the platform, Pro at $249 a month covers 50,000 searches, so searching on every turn needs Enterprise or the self-hosted SDK, where the cost becomes a gpt-5-mini call per add plus a Postgres instance.
A support operations lead at a 40-agent SaaS help desk
Let the AI assistant recall a customer's past tickets and setup
Store resolved-ticket summaries as memories keyed by customer ID and retrieve them when a new ticket opens. Keyword matching assumes English, and add-only storage means a changed plan or contact sits beside the old one, so test retrieval on real ticket histories first.
A developer building a coding agent plug-in
Persist project conventions across Claude Code or OpenCode sessions
Mem0 ships plug-ins for OpenCode, OpenClaw and Pi, and the CLI can create an API key without a dashboard. The Hobby tier's 1,000 retrievals a month suits one developer; a team moves to Starter or self-hosts.
Before you commit
- Setup
- pip install mem0ai or npm install mem0ai and set an OpenAI key; the library defaults to gpt-5-mini, text-embedding-3-small and a local Qdrant. The self-hosted server is docker compose with Postgres and pgvector, bootstrapped by make bootstrap with auth on by default.
- Integrations
- Python and TypeScript SDKs, a REST server, a CLI, and plug-ins for OpenCode, OpenClaw and Pi. Pluggable LLMs, embedders, rerankers and many vector stores, including Qdrant, pgvector and Upstash.
- Data and privacy
- Self-hosted memories stay in your stores, but every add sends conversation text to the configured LLM, and the SDK sends usage telemetry to PostHog unless MEM0_TELEMETRY=False. The platform stores memories on Mem0's servers; its pricing page cites GDPR readiness and lists SSO and on-premises only on Enterprise. SOC 2 status is not published there.
- Cost at scale
- Platform Pro is $249 a month for 500,000 adds and 50,000 retrievals; Starter is $19 for 50,000 and 5,000. Self-hosting costs a Postgres server plus one LLM extraction call per add and an embedding per search, which at gpt-5-mini prices is usually small next to the main agent's spend.
- Limits
- 863 open issues and pull requests. The April 2026 algorithm upgrade was breaking, with renamed parameters and changed add() events. Graph memory is platform-only, BM25 and entity extraction are English-only, and Docker images for the server are not tagged per release.
What's good
- Two-call API that fits around any agent loop
- Self-hosted server with auth on by default
- Open evaluation harness for its benchmark claims
- Hybrid semantic, keyword and entity retrieval
What's not
- Graph memory removed from open source in April 2026
- Headline benchmarks are platform-only
- Add-only extraction keeps outdated facts
- Telemetry on by default; keyword search English-only
Best for
Developers adding cross-session user memory to chatbots and agents.
Alternatives
Mem0: common questions
- Is Mem0 worth it?
- Mem0 is still the quickest way to give an agent a memory, but the open-source SDK is increasingly the trial version of a platform that keeps graph memory and the best benchmark scores for itself. We rate it 4.0 out of 5.
- How much does Mem0 cost?
- Open-source SDK free under Apache 2.0. Platform: Hobby free (10k adds, 1k retrievals/month); Starter $19/month; Pro $249/month (500k adds, 50k retrievals, graph memory); Enterprise custom.
- What are the downsides of Mem0?
- Graph memory removed from open source in April 2026; headline benchmarks are platform-only; add-only extraction keeps outdated facts; telemetry on by default; keyword search English-only.
- Who is Mem0 best for?
- Developers adding cross-session user memory to chatbots and agents.
- How much does Mem0 cost for a team or in production?
- Platform Pro is $249 a month for 500,000 adds and 50,000 retrievals; Starter is $19 for 50,000 and 5,000. Self-hosting costs a Postgres server plus one LLM extraction call per add and an embedding per search, which at gpt-5-mini prices is usually small next to the main agent's spend.
- Is Mem0 safe for company data?
- Self-hosted memories stay in your stores, but every add sends conversation text to the configured LLM, and the SDK sends usage telemetry to PostHog unless MEM0_TELEMETRY=False. The platform stores memories on Mem0's servers; its pricing page cites GDPR readiness and lists SSO and on-premises only on Enterprise. SOC 2 status is not published there.
- What are the best Mem0 alternatives?
- The closest alternatives we have reviewed are LlamaIndex and GraphRAG.
Answers checked Oct 23, 2026.
Make Mem0? Show this rating on your site
