Compressed visual representations reconstructed from video moments to support reasoning without storing full images.