Mitosis Labs
Research · Agent memory

Agent memory, measured.

Mitosis Cortex, evaluated on LongMemEval — the standard academic benchmark for long-term agent memory. End-to-end grading, the authors’ own judge prompts, every source cited.

500 questions~115K-token histories6 memory abilitiesICLR 2025
LongMemEvalEnd-to-end accuracy
1Mitosis Cortex
91.7%
2Supermemory
85.2%
3Zep
75.1%
4Mem0
73.8%
Cortex: preliminary, full run in progress · others: independently verified, cited below
91.7%End-to-end accuracy on LongMemEvalPreliminary · full 500-question run in progress
477 msp95 retrieval latency · 258 ms medianZep reports 632 ms · Mem0 reports 1,440 ms
100%of answers carry cited sourcesEvery retrieved fact resolves to its origin
01 — The benchmark

One benchmark counts

LongMemEval (Wu et al., ICLR 2025) is the field’s standard test of long-term memory: 500 questions over long, timestamped chat histories. Frontier assistants lose ~30 points of accuracy on it — that difficulty is what makes it worth measuring.

Questions500
History~115K tok
Abilities6
GradingLLM judge
VenueICLR 2025
LicenseMIT
01Single-session recallRetrieve a fact from one conversation.
02Multi-session reasoningStitch an answer from several conversations.
03Knowledge updatesHonor a fact that changed over time.
04Temporal reasoningReason about when things happened.
05Preference recallApply what the user has said they want.
06AbstentionKnow when the answer isn’t there.
02 — Results

Where Cortex stands

Two axes decide a production memory system: how often it is right, and how fast it answers. Every figure below is end-to-end — an LLM judge grades the final answer, never a retrieval-only proxy.

End-to-end accuracyLongMemEval · higher is better
Mitosis Cortex
91.7% prelim.
Supermemory
85.2%
Zep
75.1%
Mem0
73.8% reproduced
0255075100%
Retrieval latency, p95milliseconds · lower is better
Mitosis Cortex
477
Zep
632
Mem0
1,440
03757501,1251,500 ms

Sources: vendor publications and independent reproductions, cited in full below. Cortex accuracy is preliminary — scope stated under Method.

03 — Reading the numbers

Why most published numbers don’t compare

Recall@k is not accuracy

A “95%” headline often means the right passage appeared somewhere in the top-k results — not that the system answered correctly. Recall@k is structurally the higher number.

Self-reported scores inflate

Vendor-run harnesses reward benchmark-specific prompt tuning. Under a neutral, independent harness, one leading system’s claimed 93.4% reproduced at 73.8%.

Claimed vs. independently reproducedLongMemEval · end-to-end
Mem0
73.8% observed
93.4% claimed

A 20-point gap inside a single system. It is the reason we publish the protocol, not just the score.

04 — Method

The protocol

An evaluation we would accept from a competitor. Each step removes a way to overstate the result.

  1. 1

    Ingest unaltered

    Full multi-session histories, original timestamps, isolated workspace. No per-question tuning, no access to the answer key.

  2. 2

    Retrieve via the production path

    The same fused vector, full-text, and knowledge-graph search a live agent uses over MCP. Every retrieved item carries its source.

  3. 3

    Answer with a fixed, disclosed model

    Claude Sonnet, pinned and verified per call, answers from retrieved context alone. All settings frozen before the run.

  4. 4

    Grade with the authors’ judge

    LongMemEval’s published judge prompts, used verbatim. Judge model disclosed: Grok 4.5 (academic default is GPT-4o); cross-judge agreement is being measured.

Scope. 91.7% is a preliminary figure from a stratified sample — two questions per category — and carries a wide confidence interval. The full 500-question run is underway and will be published here under the identical protocol, regardless of outcome.

05 — Reproducibility

Check our work

The dataset is public and MIT-licensed. The judge prompts are published. We used both without modification, against the production system. A different result under the same protocol is exactly the finding this page exists to invite.

Dataset publicJudge prompts publishedJudge model disclosedRevisions dated, never silent
See Cortex in production →
Sources

References

[1]

Wu, Wang, Yu, Zhang, Chang, Yu. LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory. ICLR 2025. arXiv:2410.10813 · dataset & judge prompts

[2]

Supermemory. State-of-the-Art on LongMemEval. Vendor research page; end-to-end figures. supermemory.ai/research

[3]

Mem0. State of AI Agent Memory 2026. Claimed scores and reported latency. mem0.ai/blog

[4]

Maximem. State of AI Memory 2026: Claimed vs. Observed. Independent reproduction under a neutral harness (93.4% → 73.8%). maximem.ai/blog

[5]

Zep. Lies, Damn Lies, and Statistics. Corrected end-to-end and p95 search-latency figures; plus arXiv:2501.13956. blog.getzep.com

[6]

Liu et al. Lost in the Middle: How Language Models Use Long Contexts. TACL 2024. arXiv:2307.03172

[7]

Maharana et al. LoCoMo: Evaluating Very Long-Term Conversational Memory. ACL 2024. arXiv:2402.17753

Last updated 20 July 2026 · accuracy figure preliminary · versioned as the full evaluation completes.