Mem0
by Mem0
Persistent memory for AI agents, added in a few lines of code
Mem0 is an Apache-2.0 memory layer that sits between an AI application and its language model, extracting durable facts from conversations and recalling only the relevant ones on later turns. It gives agents continuity across sessions, users and tools, so teams stop paying to replay whole transcripts on every request, and it ships as a library, a self-hosted server or a managed cloud platform.
Mem0 is an Apache-2.0 memory layer that sits between an application and its LLM, extracting durable facts from each exchange and returning only the relevant ones on later turns instead of replaying whole transcripts. Memories are scoped hierarchically to user, session and agent, and retrieval blends semantic vector search, BM25 keyword matching and entity linking; the April 2026 algorithm moved to single-pass, ADD-only extraction with temporal reasoning, so later statements supersede earlier ones rather than accumulating contradictions. Independent coverage by InfoWorld describes the storage as a hybrid of vector, graph and key-value backends, ranked on relevance, importance and recency. In library mode it defaults to OpenAI's gpt-5-mini, text-embedding-3-small, a local Qdrant instance and SQLite history; the self-hosted Docker server swaps in Postgres with pgvector and bundles OpenAI, Anthropic and Gemini providers, while the same API is offered as a managed cloud platform with a dashboard, authentication and proprietary optimisations the vendor says open-source users will approximate but not match exactly. SDKs ship for Python and JavaScript/TypeScript, alongside an MCP integration, a CLI, an OpenMemory offering and a browser extension that shares memory across ChatGPT, Claude and Perplexity. Founded in January 2024 by Taranjeet Singh and Deshraj Yadav and based in San Francisco, the company raised $24M in total: a $3.9M seed led by Kindred Ventures and a $20M Series A led by Basis Set Ventures, with Peak XV Partners, the GitHub Fund and Y Combinator participating. Its Series A announcement reported 14 million downloads, API calls rising from 35 million in Q1 2025 to 186 million in Q3 2025, native integrations with CrewAI, Flowise and Langflow, and AWS naming Mem0 the exclusive memory provider for its Agent SDK. The repository carries roughly 63,000 stars, and the product competes directly with Zep, Letta, Cognee and LangMem.
The engineering lead who already has an agent in production and needs it to remember users across sessions without rebuilding the agent runtime — Mem0 bolts on rather than replacing the framework.
Agents recall prior context on later turns while sending far less of the transcript back to the model, cutting per-request token spend and removing the 'it forgot me' failure that kills repeat usage.
At a Glance
- Category
- Agent Development
- Pricing
- Freemium, Subscription, Usage-based
- Target Market
- CTOs, Enterprise Developers, AI/ML Engineers, Data Scientists, Startup Founders
- Deployment
- Cloud-first, Open-source, Self-hosted, API-based
- Founded
- 2024
- Headquarters
- San Francisco, United States
- Customers
- Vendor-stated 150,000+ developers; 14 million downloads and 186 million API calls in Q3 2025
Key Features
- ✓Hierarchical memory scoping
Memories are stored at user, session and agent level so a single deployment can separate what one person told the assistant from what the agent learned globally, which is what makes multi-tenant products safe to build on it.
- ✓Multi-signal retrieval
Recall combines semantic vector search, BM25 keyword matching and entity linking rather than embedding similarity alone, so exact identifiers such as an order number are found as reliably as paraphrased concepts.
- ✓Single-pass extraction with temporal reasoning
The April 2026 algorithm extracts facts in one ADD-only pass and reasons about time, so a later statement supersedes an earlier contradictory one instead of both being retrieved and confusing the model.
- ✓Identical API across library, self-hosted server and cloud
The same Python or TypeScript calls run against a local Qdrant-backed library, a Docker server on Postgres with pgvector, or the managed platform, so a prototype can move to self-hosting without a rewrite.
- ✓Provider-agnostic model and store configuration
OpenAI, Anthropic and Gemini are bundled in the self-hosted server and local models are supported, so the memory layer does not lock the application to a single LLM vendor's roadmap or pricing.
- ✓MCP integration, CLI and cross-assistant browser extension
Beyond the SDKs, Mem0 exposes memory over Model Context Protocol and through a browser extension that shares a single memory store across ChatGPT, Claude and Perplexity sessions.
Capabilities
Use Cases
- •Support agent that remembers the open ticket
A support assistant recalls the customer's product tier, previous issues and attempted fixes on the next contact, so the user is not asked to re-explain and handling time drops.
- •Cutting token spend on long conversations
Instead of resending an entire chat history each turn, the application retrieves only the facts scored as relevant, which reduces prompt size and therefore per-request model cost on long-running sessions.
- •Personalised assistant across devices and surfaces
Preferences a user states in one channel are recalled in another because memory is scoped to the user rather than the session, which is what the browser extension demonstrates across ChatGPT, Claude and Perplexity.
- •Adding memory to an existing agent framework
Teams already running CrewAI, Flowise or Langflow attach Mem0 through its native integration rather than migrating the whole agent to a different runtime just to gain persistence.
- •Self-hosted memory for data-sensitive workloads
Regulated teams run the Docker server on their own Postgres and pgvector so conversation-derived facts never leave their infrastructure, using the same API as the managed platform.
Ideal For
Best For
- ✓Adding cross-session recall to an existing LangChain, CrewAI, Flowise or Langflow agent without adopting a new agent runtime
- ✓Customer-support assistants that must remember a user's open ticket, product tier and past resolutions between conversations
- ✓Personal-assistant and companion products where retention depends on the assistant recalling stated preferences weeks later
- ✓Cutting prompt cost on long-running chat sessions by retrieving relevant facts instead of resending the full conversation history
- ✓Teams that want to prototype on the free-tier cloud API and later self-host the identical API on their own Postgres/pgvector stack
Not Ideal For
- ✗Teams that want the memory system to infer behavioural patterns and habits rather than store and retrieve stated facts — an Ask HN thread in February 2026 raised exactly this gap, and Mem0 is explicitly an extraction-and-recall layer, not a user-modelling engine
- ✗Buyers who need agent-runtime control over what enters and leaves the context window; Letta's tool-driven paging model gives that, whereas Mem0 deliberately sits outside the agent loop
- ✗Regulated procurement that requires SOC 2 Type 2 today — the published attestation is Type 1, and on-prem deployment, SSO and audit logs are all gated behind the custom-priced Enterprise tier
- ✗Small teams whose usage lands between tiers: the cloud jump from Starter at $19/month to Pro at $249/month is roughly thirteen-fold with nothing in between
Integrations
Deployment
Market & Ratings
Vendor-stated 150,000+ developers; 14 million downloads and 186 million API calls in Q3 2025
Market Analysis
Pros
- ✓Apache 2.0 with no model lock-in, and the identical API runs as a library, a self-hosted Docker server or a managed cloud service
- ✓Genuine ecosystem pull: roughly 63,000 GitHub stars, native integrations with CrewAI, Flowise and Langflow, and a well-received Show HN in September 2024 at 201 points
- ✓Lowest-friction option in its category — it attaches to an existing agent instead of demanding you rewrite onto a new runtime
- ✓Compliance posture (SOC 2 Type 1, HIPAA, GDPR, BYOK) is ahead of most memory projects, which publish nothing at all
- ✓Self-hosting on Postgres and pgvector is a first-class path, not an afterthought, so data-sensitive teams are not forced onto the cloud tier
Cons
- ✗Benchmark numbers in this category are contested and should not be trusted from any vendor's own leaderboard — competing memory projects posted to Hacker News in 2026 claim higher LoCoMo scores than Mem0 (Cortex at 73.7%, Engram at 80.0%, Forensic at 90.1%), so reproduce results on your own conversations before choosing
- ✗It stores and recalls stated facts but does not learn behavioural patterns; a February 2026 Ask HN thread was raised specifically on that gap
- ✗The vendor concedes open-source deployments only reach directionally similar results to the managed platform, because the proprietary optimisations are not in the OSS build — so self-hosting quietly costs accuracy
- ✗SOC 2 is Type 1 rather than Type 2, and SSO, audit logs and on-prem are all held behind custom-priced Enterprise
- ✗Cloud pricing has a thirteen-fold gap between the $19 Starter and the $249 Pro tier, and the retrieval counter is far tighter than the add counter, so real agent traffic exhausts a tier sooner than the headline number suggests
- ✗No verified third-party review score exists — G2 and comparable review sites return no accessible rating for this product
Pricing
Hobby
$0
- ✓10,000 add requests/month
- ✓1,000 retrieval requests/month
- ✓1 project
- ✓Community support
Starter
From $19/mo
- ✓50,000 add requests/month
- ✓5,000 retrieval requests/month
- ✓1 project
- ✓Community support
Pro
From $249/mo
- ✓500,000 add requests/month
- ✓50,000 retrieval requests/month
- ✓Unlimited projects
- ✓Private Slack support
- ✓Advanced analytics
Enterprise
Contact for pricing
- ✓Unlimited add and retrieval requests
- ✓On-prem deployment
- ✓SSO and audit logs
- ✓Custom integrations
- ✓SLA-backed support
The cloud platform is metered on two separate counters — memory add requests and retrieval requests — and the free Hobby tier gives 10,000 adds against only 1,000 retrievals, which is the limit a chatty agent hits first. Paid tiers are $19/month for 50,000 adds and 5,000 retrievals and $249/month for 500,000 and 50,000, with the vendor also offering usage-based pricing for traffic that does not map cleanly to a tier. On-prem deployment, SSO, audit logs and SLA coverage are Enterprise-only at custom pricing. The open-source library and self-hosted Docker server are free under Apache 2.0, but the vendor states self-hosted results are directionally similar rather than identical to the managed platform's, because proprietary optimisations are not included.
Security & Compliance
Connect
Sources
This page was written from 7 sources, 5 on domains other than mem0.ai.
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
Mastra
Open-source TypeScript framework and platform for building, deploying and observing production AI agents and workflows
Agentrys
Agentic design automation that lets chip teams build and own a self-improving engineering workforce
OpenAI Agents API
The Codex agent harness behind one API call — managed sessions, sandboxes and subagents
Letta
Open-source platform for stateful AI agents with persistent memory, now with a TypeScript Agents SDK for embedding them in your own apps