VoiceMem splits agent memory into an informational and an emotional brain and retrieves in 134 ms

arXiv 2608.26005, via Hugging Face Daily Papers·medium signalXBlueskyLinkedInCopy link

arXiv 2608.26005 (submitted 2026-08-26, Xie et al.) is the top-upvoted paper on Hugging Face Daily Papers this week at 167 upvotes. VoiceMem runs a parallel 'left brain' for factual memory and 'right brain' for affective attribution and dual-node persona modeling, with streaming memory I/O and interchangeable backends. At top-5 retrieval the left brain beats Mem0 at top-200 by nearly 30 points, the right brain sets state of the art on three persona benchmarks with a 4.29-point aggregate gain, and retrieval completes in 134 ms, inside standard VAD latency so it adds no conversational delay.