EmbodiedRecall: A Ring-to-Glasses System for Preserving Valuable, Fleeting Moments in Daily Activities

Abstract

Capturing meaningful everyday moments is challenging as they often occur when users are cognitively absorbed or physically occupied. While always-on recording reduces “misses,” it introduces severe privacy concerns and data triage fatigue. We present EmbodiedRecall, a ring-to-glasses system that operationalizes a proactive capture paradigm. By continuously analyzing fine-grained hand dynamics via dual IMU-equipped rings, the system recognizes compositional motion primitives and utilizes a multimodal Large Language Model (MLLM) to infer capture intent within hand-anchorable embodied moments. To balance agency with automation, EmbodiedRecall delivers subtle haptic or visual prompts, allowing users to “confirm-to-commit” short video clips from a privacy-preserving rolling buffer. A controlled lab study (N = 20) validated our multi-stage pipeline, showing that our model significantly outperforms deep learning baselines with 75.35% accuracy. A 6-day active-session deployment (N = 13) demonstrated that the system correctly aligns with user intent (46.5% confirmation rate) while surfacing serendipitous, emotionally resonant “micro-moments” users would not have captured manually. We discuss design implications for predictive wearables that enhance socio-emotional preservation while maintaining user agency.