I think most people are probably compacting conversations wrong in AI agent context management.Because the naive way is ...

I think most people are probably compacting conversations wrong in AI agent context management.Because the naive way is to tell it to read the history message by message and summarize it.From the start to the end. One chunk at a time, always updating the summarization.But you really need to understand what you're doing here. Starting from the beginning, the agent cannot really understand what is important yet. What is important is not in the historical side, it is in the current side.It should be given the context from the most recent message backwards, because then it can see what the current stuff is about, and then from the past focus on the information which is still relevant for the current stuff.Of course if you give it the whole context one-shot, then you should give it the right way around, because that is the typical order of conversations. But chunk-wise the order should be from the present to the past, so that the summary is updated with the knowledge of what is relevant *no...