Research / arXiv / Jul 16, 2026
MEMORA: Embodied Action Memory from Egocentric Videos for Reasoning and Planning
MEMORA organizes egocentric experience into editable environment, entity, activity, and inferred-knowledge stores for long-horizon robot planning.
MEMORA defines Embodied Action Memory as the ability to form, maintain, and reuse experience for later decisions. Long-horizon planning needs remembered places, object-state changes, prior procedures, and recurring patterns across actions.
The framework follows a formation, consolidation, and retrieval lifecycle. It maintains four typed stores for environment, entities, activities, and inferred knowledge, giving later planning queries structured access to past experience.
Online editing tracks identities and state histories as observations arrive. Offline consolidation turns repeated experience into reusable procedures and participant-specific regularities.
MEMORA-Bench uses 45 hours of extended EPIC-KITCHENS-100 video from 18 participants. The paper reports improvements of up to 20.5 points in memory-assessment accuracy and up to 16.6 percent in out-of-distribution Robot-Grounded Plan score against the strongest controlled baseline.
The cited arXiv record contains the memory architecture, benchmark design, language-model evaluations, qualitative robot study, and project link.
