Who this is for: teams adding long-term memory to an AI agent in 2026. Skip it if your agent only needs to remember one conversation; every major model platform now does that natively.
In short: No memory layer has yet shown, under independent test, that it makes agents better at long tasks. No independent test has yet shown a clear gain from adding Mem0 or Zep on long-horizon work. Vendor scores in the 90s have not been independently reproduced, and earlier vendor claims fell sharply when others re-ran them. So this is a list of what is worth testing, with a protocol for testing it, not a list of what works.
Key points
- No independent test has yet shown a clear gain from adding Mem0 or Zep on long-horizon work.
- Vendor scores in the 90s have not been independently reproduced, and earlier claims fell sharply when others re-ran them.
- Zep ranks first because its bi-temporal graph marks superseded facts instead of overwriting them.
- For many teams, a plain database with status, validity dates and supersession links is what wins.
- Test with 50 questions from your own history, with and without each memory layer, on the same model.
How we ranked
| Criterion | Weight | Why it matters |
|---|---|---|
| Handles change over time | 35% | The weakest category across every benchmark is "how did this change?" Memory that cannot mark a fact as superseded will blend old and new. |
| Independent evidence | 25% | Has anyone outside the vendor reproduced the claims, with a published harness? |
| Provenance and ownership | 20% | Can you see where a memory came from, export it, and self-host? |
| Adoption and backing | 20% | Funding, integrations, and whether it will still exist in two years. |
The list
1. Zep (Graphiti)
- Best for: agents whose facts change, such as accounts, positions or policies.
- Why it ranks first: it is built on a bi-temporal knowledge graph. When a new fact contradicts an old one, it marks the old fact as invalid from that point instead of overwriting it. That is the right structure for the hardest failure in agent memory.
- Watch out for: its vendor benchmark scores (94.7 and 90.2) are not independently reproduced. Graphs also cost more to keep fresh than to build.
- Pricing: free tier; paid plans are credit-based and were repriced in 2026, so check Zep's pricing page.
- What would change its rank: an independent trajectory benchmark in which the temporal graph does no better than flat memory.
2. Mem0
- Best for: teams that want the most widely adopted option and an AWS integration.
- Why it ranks here: the most traction in the category: $24 million raised by October 2025, 186 million API calls in Q3 2025, more than 13 million downloads, and built into AWS's agent SDK as a memory provider. It extracts facts into a hybrid vector-and-graph store with contradiction handling.
- Watch out for: the widest gap between vendor and independent results. Vendor scores of 92.5 and 94.4 have not been independently reproduced, and earlier claims fell sharply when others re-ran them. One 2026 benchmark put it at about one in ten correct on "how did this change" questions, whether its store was fresh or full.
- Pricing: free tier; paid plans on custom pricing.
- What would change its rank: a published, reproducible harness that closes the gap with its own claims.
3. Letta
- Best for: teams that want agent state they can inspect and control, from the research group behind the MemGPT approach.
- Why it ranks here: the OS-style "paging" design it descends from is one of the founding ideas in agent memory, and the lineage is academic rather than marketing-led.
- Watch out for: smaller and earlier: a $10 million seed at about $70 million in September 2024. Independent long-horizon results were not found.
- What would change its rank: a new funding round or independent benchmark either way.
4. Cognee
- Best for: teams already committed to a graph approach and willing to work with an early vendor.
- Why it ranks here: a memory-graph design, newly funded with a $7.5 million seed in February 2026.
- Watch out for: very early, and no independent results found.
- What would change its rank: production references in regulated firms.
5. Supermemory
- Best for: small teams and prototypes.
- Why it ranks here: a lightweight memory layer with a $2.6 million seed in October 2025.
- Watch out for: the smallest backing on the list, and no independent results found.
- What would change its rank: independent benchmark results or a larger round.
Left out, and why
- The platform-native options. Google's October 2026 agent ships four memory types, and Microsoft's Work IQ is a governed context layer. If you are already on one of those platforms, test its native memory before adding another vendor. They are a different category (bundled, not standalone), so they are not ranked here.
- A plain database with good discipline. For many teams the right answer is records with status, validity dates and supersession links in a database you already run. It is not a product, so it is not ranked, but it is often what wins.
How to test any of them in two weeks
- Build 50 questions from your own history: 20 lookups, 20 "how did this change?" questions, and 10 that assume something you have since reversed.
- Run your agent with no memory layer, then with each candidate, on the same model.
- Score correctness, and separately score how often it corrects the false premise.
- Ask each vendor for its benchmark harness, not just its score. A vendor that will not share one has told you something.
Common questions
Do AI agent memory layers like Mem0 and Zep actually improve performance?
Not yet under independent test. No independent benchmark has shown a clear gain from Mem0 or Zep on long-horizon work, and vendor scores in the 90s have not been independently reproduced.
Which agent memory layer is best for facts that change over time?
Zep (Graphiti), because its bi-temporal knowledge graph marks the old fact as invalid when a new fact contradicts it, rather than overwriting it.
How do I test an agent memory layer?
Build 50 questions from your own history: 20 lookups, 20 'how did this change?' and 10 with a reversed premise. Run with no memory and each candidate on the same model, and ask vendors for their benchmark harness.
Do I need a memory layer if my agent only needs to remember one conversation?
No. Every major model platform now handles single-conversation memory natively.
Disclosure: lowob has no commercial relationship with any vendor on this list. Facts checked 17 Oct 2026.