Beta Infinite's BetaMem paper is posted on arXiv, achieving SOTA on four memory benchmarks

Beta Infinite has written its write learning, evidence management, and self-improvement mechanisms of the BetaMem full spatiotemporal multimodal memory architecture into a paper posted on arXiv. It uses the 9B open-source model Qwen3.5-9B to generate memories and achieves SOTA on all four memory benchmarks: M3-Bench-robot, LoCoMo, OpenEQA, and EgoLifeQA. The paper decomposes memory into a temporal state graph, an append-only event log, and a source data archive of limited capacity, with state records carrying source, confidence, and version number. Ablation shows that after removing the source evidence linked to the original utterances, M3-Bench-robot accuracy drops from 48.08% to 38.61%.