Ledger preprint improves spatial recall from first-person video

Researchers introduced Ledger, a persistent three-dimensional object memory that answers spatial questions from saved records without reopening the original video. Their preprint reports accuracy rising from 29.7% to 42.6% on HD-EPIC and from 33.8% to 38.5% on UCS-Bench, with a 0.99 m median localization error on returned Ego4D predictions. Ledger links observations across recordings, retains objects after they leave view, and stores locations, histories, and short contextual descriptions. It records moves only after repeated evidence to reduce localization noise. Tests on 100 stitched streams exposed failures in both memory construction and retrieval across scene changes.