Learning to summarize user information for personalized reinforcement learning from human feedback