GROVE: Growing and Reasoning over Temporally Stratified Memory from Streaming Video Experience

Researchers introduced GROVE, a training-free memory framework for wearable video assistants.

It incrementally organizes streaming video into evidence, moments, episodes, and recurring patterns.

The same retrieval interface supports both visual question answering and proactive assistance from situational cues.

GROVE was evaluated on multiple benchmarks and showed strong results, including gains on MM-lifelong and EgoServe.