Harnesses are so wasteful and expensive

Community
1

When it comes to tool outputs - those could be referenced without remaining in context.

Active recall is thus: set an extinction rate - but allow full recall.
Should details be needed - then the model will remember the summary.
But if the reference [indirect] - then it can update the summary.
the [updated summary]
is subject to full recall, unless its summary contains a complete answer…

and… thats how you compress past tool calls without having to do them again and insert them into history.

2

Yeah, this makes a lot of sense, replaying old tool calls just to keep the full context feels pretty wasteful. A compact summary that can be expanded or updated when needed seems like a much cleaner way to handle long-running sessions.