エージェントは「記憶」でどう賢くなるのか

As models improve, agents can run for hours, but one thing stays unsolved: continuous self-learning. Because a model has only a finite context, context engineering — deciding what information to put in front of the model — matters a lot. Memory extends this across time: it is the primitive that lets a self-learning agent improve from its own experience, learning success criteria, common mistakes, and which strategies work. Claude models memory as a file system, so it can use familiar tools to keep it organized. But a single agent only sees its own task, so it misses what other agents have already learned. That is why Anthropic built dreaming — an out-of-band process that runs in the background, looks across many recent sessions, finds shared patterns and repeated mistakes, and does curation to remove duplicates and stale entries. It produces a separate, refreshed memory store that future agents can review or attach to their sessions, turning scattered memories into a more organized, high-signal knowledge base.