I like the approach; avoiding a model call just to summarize old context while keeping the cache warm is a nice property.
When it folds away an older chunk of the conversation, does it cause problems if the agent was in the middle of doing something (some multi-turn action) right when that chunk gets folded? Or is it safe to fold at any point regardless of what the agent was doing at that time?
When it folds away an older chunk of the conversation, does it cause problems if the agent was in the middle of doing something (some multi-turn action) right when that chunk gets folded? Or is it safe to fold at any point regardless of what the agent was doing at that time?