Deep sequence models tend to memorize geometrically; it is unclear why