Memory-Aware Reward Modeling: Learning Human Preferences for LLM Memory Use