We measured our memory system on the public LongMemEval-V2 Small benchmark using its own official scorer. Here is the result and everything needed to verify it — published on our terms, whether or not it is ever posted elsewhere.
Self-reported · official scorer · independent verification pendingLongMemEval-V2's headline metric, LAFS, rewards improving the accuracy-vs-latency frontier — being more accurate at a given speed. Our entry improves that frontier at the fast end: best accuracy in the sub-second regime, roughly six points above the same-speed baseline, hundreds of times faster than the high-accuracy agent baselines that take minutes per query.
| System | Accuracy | Latency | Note |
|---|---|---|---|
| RAG (query → slice + notes) | 51.0% | 0.2 s | released baseline |
| UDI v0.3.1 (ours) | 56.76% | 0.187 s | sub-second frontier |
| AgentRunbook-R | 58.6% | 26.9 s | released baseline |
| Codex (coding agent) | 69.9% | 177.2 s | released baseline |
| AgentRunbook-C | 74.9% | 108.3 s | released baseline |
At the same speed as the fastest baseline, we answer ~6 points more correctly. The high-accuracy systems win on raw accuracy but cost 100–950× the latency.
The frozen result package is content-hashed. When the package is reviewed, these digests let anyone confirm the bytes are exactly what produced the number above — nothing swapped, nothing tuned after the fact.
You don't have to trust the number — recompute it. The sealed adapter already produced its 451 answers and the official scorer already judged each one. This bundle recomputes 256/451 = 56.76% and the 0.187 s latency straight from those frozen judgments — offline, no API key, no access to the protected mechanism.
The full method, matched baselines, and negative results live in the Lab. The bundle above even prints the category breakdown — including the abstention questions, where we score lowest and say so. If the number is wrong, reproduce it and we will publish what you find, positive or negative.