A team including Sitong Gong, Caixin Kang, and Yifei Huang presents GROVE, a training-free framework that organizes streaming video into a hierarchically structured memory spanning four temporal scales: perceptual traces, time-stamped moments, coherent episodes, and cross-day patterns. Each memory stratum has a specialized retrieval skill, allowing the same unified memory to support both reactive question-answering and proactive assistance. The system reports state-of-the-art results across five benchmarks, including MM-Lifelong and EgoServe, with ablation studies showing that both the temporal organization and the scale-native retrieval interface contribute meaningfully to the performance gains.
