Running a TypeScript codebase through a long AI session: by message 40 something always slips. Not hallucinations. The model loses track of decisions already made. I've started summarizing key decisions every 15 messages.

Does that match your experience, or is there a better pattern?