Mem0
Also known as: mem0, mem-zero, mem0ai
Open source memory layer that gives AI agents persistent, personalized recall across sessions through a simple add and search API, cutting token costs.
Mem0, pronounced mem zero, is a memory layer for AI agents and assistants that addresses a basic limitation of large language models: they are stateless, so every call starts from a blank slate, which makes agents forget across sessions and forces teams to replay entire conversation histories on each turn. After each interaction, Mem0 extracts the durable facts, a user's preferences, decisions, and prior context, stores them, and on the next call retrieves only the relevant ones to inject into the prompt. The result is agents that remember across sessions and meaningfully lower token costs, since a handful of relevant memories replace the full transcript. It is open source under Apache 2.0 with roughly forty eight thousand GitHub stars.
Mem0 exposes its capability through a simple add and search API and is framework agnostic, working with LangChain, LangGraph, CrewAI, and AWS Strands rather than being its own agent framework. It is not a database itself; it orchestrates a vector store such as Qdrant for semantic recall alongside Postgres for history, and offers a graph memory extension that captures entity relationships for richer temporal reasoning. It supports user, session, and agent level memory scopes so one agent can keep separate context per user, and handles adaptive updates that revise an existing memory when a user corrects a preference rather than creating duplicates, plus conflict resolution and decay of outdated facts.
There are three ways to run it. The open source library suits testing, self hosting suits teams that want to run Mem0 alongside their own vector store and Postgres, and the managed Mem0 Platform offers zero operations cloud. The company publishes an open source evaluation framework and reports strong results on long term memory benchmarks like LoCoMo and LongMemEval. An unusual touch is agent native onboarding, where an AI agent can mint a working API key from the command line in seconds without a dashboard or email.
For enterprises, Mem0 adds governance, full read and write audit logging, SOC 2 and HIPAA compliance, bring your own key, and deployment on Kubernetes, private cloud, or air gapped environments. The managed platform meters memory add and retrieval requests. A free Hobby tier includes 10,000 add and 1,000 retrieval requests a month, Starter is $19 a month, Growth is $79 a month with more requests and projects, and Pro is $249 a month with graph memory and unlimited projects. A custom Enterprise tier adds on premises deployment, audit logs, SSO, and an SLA, and usage based pricing is available.
Vendor details
Canonical URL
https://mem0.ai
Category
Agent infrastructure
Subcategory
Agent memory
Funding status
Founded by Taranjeet Singh (CEO) and Deshraj Yadav (CTO), the team behind Embedchain, with around $24M raised. The open source project has roughly 48,000 GitHub stars. Independent.
Company status
independent
Use cases & customers
Primary use cases
Target customers
Deployment options
Integrations
Framework agnostic with a simple add and search API and SDKs, integrating with LangChain, LangGraph, CrewAI, and AWS Strands Agents, and exposable as an MCP server. Self hosting orchestrates a vector store like Qdrant plus Postgres, and managed backends include Amazon ElastiCache for Valkey and Neptune Analytics for graph memory.
In practice
Your support bot asks returning users the same questions every session. You add Mem0 keyed by user, so it stores each user's preferences and recalls them next time without replaying the whole transcript.
Your token bill is high because you stuff full chat history into every prompt. Mem0 extracts and retrieves only the relevant memories, cutting context size and cost while keeping the agent's recall.
You need agent memory that stays in your environment for compliance. You self host Mem0 alongside your own Qdrant and Postgres, or deploy it air gapped, keeping every stored memory inside your boundary.
Sources & related URLs
Related / legacy domains
Research sources
Capability coverage
5.5 / 14 capabilities · 39%
| Integrations & Tool CallingFramework agnostic integrations with LangChain, LangGraph, CrewAI, and AWS Strands, plus exposure as an MCP server that agents call to read and write memory, but it is a memory layer rather than a tool calling hub. | Partial |
|---|---|
| Workflow OrchestrationProvides a memory layer that agent frameworks call, but does not orchestrate agent workflows, sequencing, or branching itself. | Unable to verify |
| Knowledge Grounding & RAGProvides semantic, retrieval augmented recall of stored memories using embeddings and a vector store, RAG like in mechanics, though it grounds on conversation derived memories rather than a document knowledge base. | Partial |
| Human Oversight & GuardrailsOffers no human review, approval workflow, or content guardrails; it stores and retrieves memories. | Unable to verify |
| Security, Identity & GovernanceSOC 2 and HIPAA compliance, bring your own key, zero trust, full read and write audit logging, and SSO on Enterprise, a solid governance posture for a memory layer, though some controls are Enterprise gated. | Partial |
| Observability & AuditabilityFull read and write audit logging records what, who, and when, and teams can inspect which memories are retrieved for a query to explain agent behavior, plus analytics on higher tiers, though it is memory observability rather than full agent tracing. | Partial |
| Memory & State PersistenceCore product. A dedicated persistent memory layer with user, session, and agent level scopes, episodic, semantic, procedural, and associative memory types, adaptive updates and conflict resolution, decay of outdated facts, and a graph memory extension. | Full |
| Deployment & Data ResidencyOpen source under Apache 2.0 and fully self hostable alongside your own vector store and Postgres, with managed cloud, on premises, private cloud, Kubernetes, and air gapped deployment plus bring your own key, giving full deployment and data residency control. | Full |
| Prebuilt Agents, Templates & PacksProvides example integrations but no prebuilt agents, templates, or installable packs. | Unable to verify |
| Triggers & Channel CoverageExposed as an API that agents call; provides no triggers, scheduling, or conversational channel coverage of its own. | Unable to verify |
| Model Flexibility & RoutingProvider and model agnostic, working with any LLM for the memory extraction step and avoiding model lock in, but it is not a model routing gateway. | Partial |
| APIs, SDKs & MCP ExtensibilityA clean add and search API, Python and TypeScript SDKs, a CLI, a fully open source codebase, and exposure as an MCP server give comprehensive extensibility. | Full |
| Testing, Debugging & OptimizationPublishes an open source evaluation framework to reproduce its own memory benchmarks, but offers no user facing agent testing, debugging, or optimization product. | Unable to verify |
| Browser & Computer UseNot applicable. Mem0 is a memory layer and does not provide browser automation or computer use; its browser extension only captures memories. | Unable to verify |
Pricing
From $19/mo · free tier (10k add, 1k retrieval/mo) + open source
Subscription tiers metered by memory add and retrieval requests per month and number of projects; usage based pricing also available
Included quota
Hobby free (10k add, 1k retrieval/mo, 1 project). Starter $19/mo (50k add, 5k retrieval, 1 project). Growth $79/mo (200k add, 20k retrieval, 3 projects, email support, basic analytics). Pro $249/mo (500k add, 50k retrieval, unlimited projects, private Slack, advanced analytics, graph memory). Enterprise custom (unlimited).
What is public
Full self serve tier pricing (Hobby free, Starter $19, Growth $79, Pro $249) with add, retrieval, and project limits is public; Enterprise is custom and usage based pricing is available.
Billing mechanics
Subscription tiers metered by monthly add (write) and retrieval (read) requests and project count. Each add request triggers an LLM call to extract memories. Open source self hosting has no request metering; enterprise adds on-prem and SLA under custom pricing.
Cost watchouts
Each memory add triggers an LLM call you pay for, so high turn chat workloads inflate both request counts and underlying model cost; self hosting moves spend to your own vector store and Postgres infrastructure.
Variable cost rationale
Cost scales with the volume of memory add and retrieval requests, which grows with conversation and agent traffic, and add operations involve an LLM call to extract facts, so high write workloads consume quota and cost faster. Self hosting shifts cost to your own infrastructure.
Additional watchouts
Add requests are metered and each involves an LLM call, so high write chatbots consume quota and cost quickly; batching or downsampling writes is recommended. Graph memory is Pro and above, and on-prem, audit logs, and SSO are Enterprise only.
Overage / add-ons
Plans cap monthly add and retrieval requests by tier; exceeding a tier's quota means upgrading, and the company also offers usage based pricing for needs that do not fit a tier. Self hosting the open source build carries no request metering.
Sales call required
No — self-serve available
Free / trial
Free Hobby tier: 10,000 add and 1,000 retrieval requests/month, 1 project, community support. Open source (Apache 2.0) and self hostable at no license cost.
Lowest paid plan
Starter $19/mo (50k add, 5k retrieval requests/mo)
Commercial notes
Open source led, developer first adoption with a free tier and a free Apache 2.0 self host path, scaling to managed tiers and an Enterprise plan for on-prem, SSO, and audit logging. Strong open source traction at roughly 48k stars.
Key ambiguities
Enterprise dollar pricing and usage based per request rates above plan quotas are not public.
Cancellation / refund
Hobby, Starter, Growth, and Pro are self serve subscriptions with standard cancellation. Enterprise terms are contractual.
Support SLA / resale
Community support on Hobby and Starter, email on Growth, private Slack on Pro, and private Slack with an SLA on Enterprise.
Missing data
Enterprise dollar pricing and the exact usage based per request rates are not public.
Related vendors
- AgentOps — Agent observability and reliability platform with broad model and…
- Agno — High-performance agent runtime and framework (formerly Phidata) with…
- Apify — Cloud platform for web scraping and automation with 45,000+ prebuilt…
- Arcade — Authenticated tool calling platform and MCP runtime that handles…
- Arize AI — AI observability and evaluation platform that traces, evaluates, and…
- Braintrust — AI evaluation and observability platform with self-serve pricing,…