Most AI chatbots forget everything the moment a conversation ends — every session starts from zero, even if the user explained their situation in detail yesterday. "Agent memory" tools fix this by storing what a user has told an agent and retrieving the right pieces later, instead of stuffing an entire chat history back into the prompt every time. Two of the most-discussed options in this space are Mem0, a generalist memory API, and Zep, which specializes in tracking how facts change over time. They overlap enough to compete directly, but they're built around different bets.
Mem0
Mem0 gives AI chatbots and agents persistent memory through a single API that combines vector, graph and episodic memory in one product. It's the most widely adopted tool in its category — over 61,000 GitHub stars — and it's the memory provider baked into the AWS Agent SDK, which is about as strong a third-party validation as this space gets.
Pricing: Freemium — Hobby is free for 10K memory-add requests/month, Starter is $19/month for 50K requests, Pro is $249/month for 500K requests with unlimited projects and graph memory, and Enterprise covers on-prem/SSO/SLA on request.
Strengths: the default memory provider for AWS's Agent SDK; by far the largest community and adoption in the category; combines vector, graph and episodic memory instead of forcing a choice between them; SOC 2 Type 1 and HIPAA compliant.
Limits: it's a generalist rather than specialized on any one memory approach, so teams with a narrow, specific need (like Zep's temporal focus or Letta's self-hosted control) may find a specialist tool fits better; Pro-tier costs need checking against real traffic for heavy graph-memory users.
Zep
Zep builds a temporal knowledge graph — a living map of what a user has told an agent and how those facts connect and change over time — so the agent can pull the exactly-right context in under 200ms instead of re-reading an entire conversation history. The temporal angle is the differentiator: Zep is built to handle facts that get updated or contradicted, not just facts that get added.
Pricing: Freemium — Free tier gives 10,000 credits/month across 2 projects. Flex is $104/month (billed annually) for 50,000 credits and 5 projects. Flex Plus is $312/month (annual) for 200,000 credits and 10 projects. Enterprise is quoted on credits and SLA.
Strengths: a temporal knowledge graph that captures how facts evolve, not just static similarity; claims sub-200ms context retrieval at scale; SOC 2 Type II and HIPAA BAA certified, which matters for regulated-industry deployments.
Limits: more specialized and less adopted (under 5,000 GitHub stars) than a generalist like Mem0; the temporal-graph approach can be more complexity than a simple memory need requires; the free tier's 10,000 credits get used up fast in serious testing.
Side-by-side
| Mem0 | Zep | |
|---|---|---|
| Approach | Generalist: vector + graph + episodic memory | Specialist: temporal knowledge graph |
| Free tier | 10K add-requests/month | 10K credits/month, 2 projects |
| Entry paid tier | $19/month (50K requests) | $104/month billed annually (50K credits) |
| Compliance | SOC 2 Type 1, HIPAA | SOC 2 Type II, HIPAA BAA |
| Adoption | 61,000+ GitHub stars, AWS Agent SDK default | Under 5,000 GitHub stars, more niche |
| Best for | Most teams adding memory to an agent or chatbot | Regulated teams needing facts-over-time reasoning |
Other options worth checking
Both tools list the same two alternatives worth a look. Letta takes a third approach entirely: an open-source, Apache 2.0 framework built on academic MemGPT research that you self-host and fully control, rather than calling a hosted API — more setup work, but no dependency on a third party's infrastructure. Pinecone is a different category of tool — a pure vector database rather than a purpose-built memory layer — worth considering if you want to build your own memory logic on top of fast vector search instead of adopting an opinionated memory product.
Verdict
Pick Mem0 if: you want the most broadly adopted, generalist memory layer, you're already in the AWS ecosystem, or you want vector, graph and episodic memory available without picking just one approach.
Pick Zep if: your agent needs to reason about facts that change over time — not just recall static facts — especially in a regulated industry where SOC 2 Type II and HIPAA BAA matter.
For most teams adding memory to a first agent, Mem0's generalist approach and larger community make it the safer default. Reach for Zep specifically when the temporal angle — tracking what changed, not just what was said — is central to what you're building.