Compaction works much better than many implicitly predicted via emphasizing continual learning as a huge blocker

@Miles_Brundage Funny I am currently putting the finishing touches on an article more or less arguing the opposite.

@Miles_Brundage I think it's both true that compaction has worked better than I would have predicted a couple of years ago and also that it's far from adequate for solving continual learning well enough to get to "AI scientists" or whatever.

@binarybits @Miles_Brundage Somewhere in between Claude Opus 4.5 and 4.7 context management went from terrible to really good. It is becoming clear this is not a show-stopper and hierarchical memory management is the way to go, rather than retraining or new architecture.