I've been doing some research on this question: At inference time a large part of a model's working memory lives in the KV cache, plus whatever external memory the harness bolts on. I've been poking at the storage-and-retrieval side of this, treating that cache as an index, and what stands out is th
