Research
GROVE Builds One Temporally Stratified Memory From Streaming Video, Serving Both Reactive Q&A and Proactive Assistance
Most video-memory systems support only question-conditioned recall while proactive assistants bolt on separate memory and control mechanisms. GROVE is a training-free framework that grows a single memory causally from a continuous video stream, retaining fine-grained perceptual evidence and consolidating it into time-stamped moments, coherent episodes, and recurring cross-day patterns, each stratum paired with a scale-native retrieval skill. It reports best results among compared methods on MM-lifelong and EgoServe, with ablations showing the pattern stratum contributes most when evidence spans multiple days; code is promised at github.com/SitongGong/GROVE.
↳ Follow the thread