“Cursor uses "cache warming" to reduce latency by pre-loading the KV cache with context, like the current file's contents, as a user types a prompt.”