I am also trying to understand the difference between compaction, and what IDEs like Cursor do when they "summarize" context over long-running conversations.

Is this saying that said summarization now happens at the model level? Or are there other differences?