Asked about Working with inputs bigger than you want to send

What would you keep when compressing a long conversation?

An answer that works

The goal, ids and numbers, constraints the user stated, decisions made and attempts that failed, ideally in a structured record, plus the most recent turns word for word.

Reading an answer is not being able to give one. The section that teaches this is Working with inputs bigger than you want to send, 4 hours of Context windows.

Learn this properly

Also asked about Context windows

What counts against the context window?Is recall uniform across a long context window?Why does a long conversation get expensive faster than it gets long?A request does not fit. What do you cut first?A model supports a very long context. Why not always send everything?