How do Agents avoid context drift?

My experience with Claude chats is that if a conversation exceeds 10-20 back and forth char inputs and responses that thighs go down hill and old errors creep in and instructions get set aside. It's time to write a checkpoint and restart a new chat. Presumably persistent agents have very long interaction histories. So how does an agent not drift. To be concrete say an agent that summarizes your emails and books calendar events and so forth. It tries to learn your habits and preference over the years. How do