Research
We build long-term memory for AI. Our research covers how models remember, recall, and stay continuous over time—on public benchmarks, with every run reported.
Tablet 2 scores 95.7% on LongMemEval-S and 67.5% on BEAM-1M, every run published, and it finds a photograph stored with no caption from a sentence in any language, where keyword search has nothing to score at all. Most of the paper is about how little a score means on its own, and it reports three findings that run against us at the same weight as the rest.
Read the paper →ALL PUBLICATIONS
Open source
Every number we publish can be reproduced. The tools we use to measure, and the SDKs you write against, are open here.
Benchmark glasshouse
Benchmark results, published with everything needed to recompute them.
beam1m-tablet-2
The harness, the scoring scripts and the per-question records behind every number in the tablet-2 paper.
The questions stay with the benchmark authors. You get those from them and join.