I haven't read it yet to be certain, but my impression is that rather than using lossless compression of entire chats, this uses lossy summarization to get the

I haven't read it yet to be certain, but my impression is that rather than using lossless compression of entire chats, this uses lossy summarization to get the gist of chats. There will be tradeoffs between the two methods, hopefully this paper covers that.