Evidence
We do not top numbers. Every figure here comes with its method, its source and a way to verify it yourself.
Last updated: 7 September 2026 · every figure checked that day against arXiv 2604.23878 and the npm registry
A benchmark number is only worth the method behind it. Some vendors report higher figures — up to “95.4 %” — on the basis of their own, undisclosed procedures. We take the opposite route: fewer superlatives, and for every number the method, the citable source and a repro or live link.
Three of nine direct comparisons hold, six are ties, none is lost — three competitors, judged by three independent LLM judges.
| Figure | What it measures | Method | Source & verification |
|---|---|---|---|
| 47.7 %vs. 31–43 % comparison methods | Share of questions answered correctly across long, distributed conversation histories (LongMemEval-500). | Scored by the official LongMemEval judge (gpt-4o-mini, correct or incorrect per question), not by us. | Preprint · Methodology → |
| 3 · 6 · 0hold · ties · losses | Direct pairwise comparisons against established methods. | Three independent LLM judges, version-matched: three hold against A-Mem, six are ties against Letta and Mem0, none lost. | Preprint → |
| 91.3 %at 1/109.6 of the token budget | Long-context oracle accuracy at a fraction of the tokens. | Measured against a full-context oracle; details in the preprint. | Analysis article → |
| 159 foundational + 6 PMA components | Mechanisms in the memory core, each ablated and measured separately. | Nine foundational algorithms open-sourced; the six PMA components documented. | Open Source · Technology → |
| Apache 2.0open core on npm | The memory core is openly licensed and installable. | Public repository, npm package with provenance, no external dependencies. | Run it live → |
| Open AccessarXiv preprint + DOI | The full paper including replication material. | Preprint on arXiv (2604.23878), permanently archived with a DOI on Zenodo. | Publications → |
All empirical results come from the open-access preprint (arXiv 2604.23878); the core benchmark is LongMemEval-500, scored by three independent LLM judges. Full methodology, assumptions and limitations are on the methodology and publications pages.
The figures above are about ZenBrain alone. To read them against individual systems, there are three pairwise comparisons here and one overview of nine open systems that deliberately awards no ranking.
More from this research
Core and application fields, on a shared ethics foundation.