How the KV cache gives every AI conversation a physical weight in silicon, and what happens when the memory runs out.