← Back to Computer Vision
cs.CV

Can AI glasses remember what you did yesterday?

Samiul Alam, Shakhrul Iman Siam, Michael J. Proulx, James Fort, Richard Newcombe, Hyo Jin Kim, Mi Zhang

May 30, 2026

Existing video AI excels at recognizing actions in short clips, but falls apart on realistic memory tasks over long egocentric streams. SuperMemory-VQA introduces 4,853 grounded questions from 52.9 hours of everyday activities recorded with AI glasses, spanning object/location memory, intent recall, and timeline reconstruction. Benchmarking shows leading LLMs struggle significantly—the field needs architectures that admit uncertainty rather than hallucinate answers.
Published as SuperMemory-VQA: An Egocentric Visual Question-Answering Benchmark for Long-Horizon Memory arXiv:2606.00825
Read the original paper →