Capture Before the Question: A Natural-Image Visual Retention Pilot

A visual assistant may store a scene before learning which detail a user will request. This motivates evaluating retained evidence with future questions rather than assuming that pixel fidelity or detected text predicts answer quality. We report a frozen, small natural-image pilot using TextVQA's published human answers. Four development and 24 image-disjoint held-out groups are evaluated with two local vision-language models, for 224 actual calls. Question-blind JPEG, WebP and OCR-crop-plus-thumbnail representations share a 12,000-byte cap that includes their ZIP containers and metadata; the original image is an unbudgeted reference. On held-out images, official human-consensus mean scores for JPEG/WebP/OCR crops are 0.700/0.771/0.538 with Qwen3.5 and 0.733/0.767/0.450 with Gemma4, versus 0.883/0.779 for originals. The crop baseline never outperforms WebP on a paired held-out image for either model. All 84 retained packets fit the cap. One original-image response violates the required string-answer contract and remains scored zero without repair. These observations support ordinary-codec baselines and direct utility measurement. They establish neither a better encoder nor energy savings. We outline a larger proxy-reliability study and a lightweight capture-time codec controller, with learned representation development conditional on a demonstrated gap.