One of the largest AI companies on Earth cannot figure out an algorithm for when not to drop caches in long-running sessions?