Give an AI chatbot or agent a way to remember things across conversations — instead of forgetting everything the moment a chat window closes — and product quality jumps immediately: it stops re-asking questions you already answered and starts feeling like it actually knows you. Letta and Mem0 are two of the most-adopted ways to add that memory layer, but they take almost opposite approaches: one hands you a framework to self-host, the other hands you a hosted API.
Letta
Letta is an open-source framework for building AI agents that keep their state and memory under your own control — self-hostable, and built on real academic research (MemGPT, from UC Berkeley) rather than a purely commercial pitch.
Pricing: Free tier gives you 3 stateful agents, bring your own API keys. Personal Pro is $20/month (20 agents, remote sandbox). The Developer API is $20/month plus $0.10 per active agent/month plus LLM usage costs. Teams Pro is $20/seat/month. Enterprise is custom.
Strengths: genuine academic roots rather than pure marketing, fully open source (Apache 2.0) with complete self-hosting possible, 24,000+ GitHub stars, and Letta Code topped the independent Terminal-Bench benchmark.
Limits: it's a framework you build on rather than a plug-and-play API like Mem0, meaning more setup work in exchange for more control, and the per-active-agent pricing needs careful modeling at scale.
Mem0
Mem0 gives AI chatbots and agents persistent memory through a straightforward API — so they remember what you told them last week instead of forgetting the moment a conversation ends, which is the default for most AI apps.
Pricing: Hobby tier is free for 10,000 memory-add requests/month. Starter is $19/month for 50,000 requests. Pro is $249/month for 500,000 requests, unlimited projects and graph memory. Enterprise is custom, with on-prem, SSO and SLA options.
Strengths: it's the exclusive memory provider for the AWS Agent SDK — strong third-party validation — the most-adopted project in its category at 61,000+ GitHub stars, it combines vector, graph and episodic memory in one product, and it's SOC 2 Type 1 and HIPAA compliant.
Limits: it's a generalist rather than specialized on one memory approach (unlike Zep's temporal focus or Letta's full self-hosted control), and heavy graph-memory users should check whether the Pro tier's limits fit their real traffic before committing.
Side-by-side
| Letta | Mem0 | |
|---|---|---|
| Model | Self-hosted framework | Hosted API |
| Free tier | 3 stateful agents | 10,000 requests/month |
| Entry paid tier | $20/month | $19/month |
| GitHub stars | 24,000+ | 61,000+ |
| Compliance | Not highlighted | SOC 2 Type 1, HIPAA |
| Best known for | Full control, academic roots | Broadest adoption, plug-and-play |
A third option: Zep
Zep takes yet another angle: a temporal knowledge graph that tracks how facts change over time, not just what was said, with sub-200ms context retrieval claimed at scale. Pricing starts free (10,000 credits/month, 2 projects) and moves to $104/month for the Flex tier. It's SOC 2 Type II and HIPAA BAA certified, which makes it a fit for regulated industries — though it's less adopted (4,770 stars) than Mem0, and its free tier runs out fast in serious testing.
Verdict
Pick Mem0 if: you want the fastest path to adding memory via API, the broadest community and battle-testing, and don't need to self-host.
Pick Letta if: you want full control over where agent state lives, you're comfortable building on a framework rather than calling an API, and open-source transparency matters to your stack.
Consider Zep if: you're in a regulated industry and need SOC 2 Type II / HIPAA compliance plus a memory model built specifically around facts that change over time.