Comparing AI memory layers
Buyers ask how Anamnesis differs from Mem0 or Zep often enough that the answer deserves its own page. Here it is, with a warning attached: the one thing you cannot do with published memory benchmarks is line the scores up in a column and pick the biggest number.
The comparison that is fair to make
Prices are hosted entry tiers as of August 2026. Benchmark figures are each vendor's own published claim, under their own protocol. They are not comparable to each other, for reasons the next section explains.
| Anamnesis | Mem0 | Zep | Letta | Supermemory | |
|---|---|---|---|---|---|
| Core architecture | Episodes distilled to gists and dated propositions, with provenance tiers and belief revision | Fact extraction into vector and graph stores | Temporal knowledge graph (Graphiti) | Agent runtime with in-context memory management (MemGPT lineage) | Universal memory API over user documents |
| Open source | Local mode, free | Apache-2.0 core | Graphiti engine | Yes | Yes (core) |
| Hosted entry price | $19/mo planned, free alpha today | $19/mo Starter | $125/mo Flex | $20/mo Pro | roughly $7 to 19/mo |
| Published benchmark claim | 77% LongMemEval-S, locked holdout, two judges from different model families | 92.5 on LoCoMo, self-scored | 71.2% on LongMemEval, own paper, gpt-4o judge | Paper evaluations (MemGPT) | None published |
| Cross-product memory | One corpus across Claude, ChatGPT, Codex, and Gemini | SDK integrations | Agent frameworks | Own agent stack | Consumer apps plus API |
Each of these products is good at something real. Zep's temporal graph is serious infrastructure for enterprise agent fleets. Mem0 has the broadest SDK ecosystem in the category. Letta is as much an agent runtime as a memory layer, and if you want the agent too, that matters. We built Anamnesis for a different job: one person's working memory, carried across every assistant they use, with an audit trail for how each fact got there and what superseded it.
Why the numbers do not line up
Mem0 reports 92.5 on LoCoMo. Zep's paper reports 71.2% on LongMemEval. Our holdout run scored 77% on LongMemEval-S. Three different benchmarks, question sets, judge models, and scoring prompts. Ranking these against each other is meaningless, and any vendor who does it in a sales deck is hoping you will not check.
The cautionary tale is public. Zep published an 84% LoCoMo claim; an independent re-evaluation with corrected settings measured 58.44% on the same benchmark. A 26-point swing on identical questions, from evaluation choices alone: which judge, which prompt, which questions get excluded, and how many times the team iterated against the test set before reporting.
That last one is the quiet killer. If you tune on the questions you report, the score measures your tuning, not your memory. Our protocol locks the holdout before the run, forbids iteration on it, burns any question we ever debugged against, and requires two judges from different model families to agree before an answer counts. The full protocol, prompts included, is published here.
So we make the narrower claim: not that 77 beats 71, but that our number survives a protocol you can audit. If you are evaluating any memory vendor, us included, ask for four things: a holdout locked before the run, a judge from a different model family than the system under test, published scoring prompts, and a policy for burning contaminated questions. The answers are informative either way.
Where to go deeper
The benchmarks overview has the full LongMemEval results and protocol. The FAQ answers the RAG question. For the sovereignty side, hosted to fully on-premises, see the sovereignty ladder.