Memorizon proposes minutes-long video training for world models

A researcher introducing Memorizon says world models trained on roughly 10-second clips rarely see both a first visit to a place and a return, limiting what they can learn about consistency. Memorizon trains on minutes-long videos, the researcher says, while each chunk attends only to a small bank of retrieved frames; they say the cost stays bounded.