Mem0

What is Mem0?

Mem0 is a memory layer that gives AI agents and chat applications persistent context across sessions. You send it conversation data, it extracts compact memories such as preferences and facts, and later it retrieves the most relevant ones for each new prompt. Memories are stored at user, session and agent level. The company says this cuts redundant context, lowers token costs and speeds up responses compared with replaying a full chat history every time.

Taranjeet Singh, the chief executive, and Deshraj Yadav founded Mem0, a Y Combinator company. In October 2025 it announced $24 million across seed and Series A rounds led by Kindred Ventures and Basis Set Ventures, with Peak XV Partners and GitHub Fund also investing. The core code is open source under Apache 2.0 with more than 60,000 GitHub stars, and a managed platform adds hosting, analytics and enterprise controls. Mem0 says over 150,000 developers build with it.

Key Features

• Memory extraction and compression: Mem0 condenses chat history into short memories instead of storing every message. Its compression engine keeps the useful context while trimming tokens, so prompts stay smaller and responses arrive faster.
• Multi-level memory: Mem0 organises memory into user, session and agent levels. User memory holds long-term preferences, session memory tracks the current conversation, and agent memory records system and action state for multi-step workflows.
• April 2026 memory algorithm: Mem0’s April 2026 algorithm scored 92.5 on LoCoMo and 94.4 on LongMemEval, according to its GitHub repository. It relies on single-pass extraction, entity linking, multi-signal retrieval and temporal reasoning, using roughly 7,000 tokens per query.
• Graph memory: Graph memory links entities such as people, projects and places across stored memories. It helps agents answer relationship questions, although the managed platform includes it only on the Pro plan and above.
• Dream memory consolidation: Dream is a Pro feature that consolidates stored memories over time. It merges and tidies related facts, so the memory store does not fill up with duplicates as conversations grow.
• Three deployment options: Mem0 runs as a Python or Node.js library, a self-hosted Docker server or a managed cloud platform. Teams can prototype locally for free, then move to hosted infrastructure without rewriting their memory logic.
• Developer tooling: Mem0 ships SDKs for Python and Node.js, plus a REST API, a CLI and an MCP server. The MCP route lets coding assistants and agent clients read and write memories without custom integration code.
• Enterprise security: Mem0 states it is SOC 2 Type I certified, HIPAA compliant and GDPR ready. Enterprise plans add on-premises deployment, audit logs, SSO and custom integrations, and the platform supports bring-your-own-key encryption.

Pros

• Open-source core: Mem0’s core library is open source under the Apache 2.0 licence, so you can inspect, modify and self-host it. That reduces lock-in compared with memory services sold only as closed APIs.
• Generous free tier: The free Hobby plan includes 10,000 add requests and 1,000 retrieval requests a month with unlimited end users. One directory notes that no credit card is needed, which keeps early prototypes cheap.
• Strong benchmark results: Mem0 reports scores of 92.5 on LoCoMo and 94.4 on LongMemEval after its April 2026 update. Those company figures point to accurate recall over long conversations, though you should still test against your own data.
• Wide framework support: Mem0 plugs into many agent frameworks, including LangGraph, CrewAI, LangChain, AutoGen, LlamaIndex and the Vercel AI SDK. One guide counts 21 supported frameworks, so most stacks need little glue code.
• Model and database choice: Mem0 works with OpenAI, Anthropic, Google, AWS Bedrock, Groq, Mistral and local models through Ollama, according to one directory. It also supports more than 20 vector databases, including Qdrant, Pinecone and Weaviate.
• Proven adoption: Mem0 reports over 150,000 developers and lists firms such as AWS, Nvidia and PwC among its users. AWS chose it as the exclusive memory provider for its Agent SDK, according to coverage of the 2025 funding round.

Cons

• Steep jump to Pro: Mem0’s paid plans leap from $19 a month for Starter to $249 for Pro, with no tier in between. Growing apps that outgrow 5,000 monthly retrievals therefore face a large cost increase.
• Graph memory gated: Graph memory and Dream consolidation are limited to Pro and Enterprise on the managed platform. Starter users get vector-only retrieval, which weakens answers that depend on relationships between entities.
• Hobby data use: One directory reports that content from Hobby plan users may be used to train Mem0’s models. Teams handling customer data should read the terms carefully or move to a paid or self-hosted setup.
• Compliance limits: Mem0 holds SOC 2 Type I rather than Type II, and one review describes its HIPAA position as self-attested. Large regulated buyers may need extra assurance before procurement signs off.
• Extraction latency and cost: Mem0 uses an LLM to extract memories, so each add request carries model cost and processing time. Its GitHub benchmarks show median latency of about 0.9 to 1.1 seconds, which adds up in real-time chat.

Who is Using Mem0?

• AI application developers: Builders add persistent user preferences to chatbots, copilots and personal assistants.
• Agent framework teams: Teams on LangGraph, CrewAI and similar frameworks give multi-step agents long-term state.
• Enterprises: Companies such as PwC, Nvidia and Samsung appear among the users Mem0 lists on its site.
• Cloud platform builders: AWS uses Mem0 as the memory provider for its Agent SDK, as reported in funding coverage.

Pricing

Mem0 is freemium, with a free Hobby tier and paid plans at $19 and $249 a month (official prices, October 2026).

Mem0 also offers usage-based pricing for traffic that does not fit a fixed tier and mentions a startup programme. Self-hosting the open-source version removes platform fees, but you then pay for your own LLM calls, embeddings and vector database, so compare both routes before committing.

What Makes Mem0 Unique?

An open-source library, a managed platform and published research from the same team let you start locally and scale without switching vendors. Mem0 stores extracted, compressed facts rather than raw chat logs, which keeps prompts small. The company publishes results on LoCoMo, LongMemEval and BEAM, and its role as the memory provider for the AWS Agent SDK gives it unusual reach for a startup founded so recently.

Compatibilities and Integrations

• SDKs and APIs: Python and Node.js SDKs, a REST API, a CLI and an MCP server.
• Frameworks: LangGraph, CrewAI, LangChain, AutoGen, LlamaIndex and the Vercel AI SDK.
• Models and storage: OpenAI, Anthropic, Google, Bedrock, Groq, Mistral and Ollama, plus more than 20 vector stores.
• Deployment: Managed cloud, self-hosted Docker server, or on-premises on the Enterprise plan.

Mem0 Tutorials

A basic memory loop needs only a few lines of code.

1. Install the library with pip install mem0ai or npm install mem0ai.
2. Create a free Hobby account and copy your API key, or run the self-hosted server.
3. Add a conversation for a user ID so Mem0 extracts memories.
4. Search that user’s memories before each new prompt.
5. Pass the retrieved memories into your model’s context and review them in the dashboard.

How We Rated It

Summary

Mem0 suits developers who want agents and chatbots to remember users without building their own memory pipeline. Its main limitation is the price gap between Starter and Pro, which also gates graph memory. Self-hosting the open-source version softens that cost. Start on the free Hobby plan with test data, then measure retrieval quality on your own conversations before choosing a tier.