Desktop app feature request - Memory/history/context aware AI assistant

[zedd3v]

I cannot tell if my model context retrieval is poor or if magic context is overpruning.
Would be nice to have an AI assistant we could ask these questions to, e.g "Was X context pruned?".
And possibly other questions regarding memories / history and even global context retrieval for project related questions.

[ualtinok]

I'm not sure a full fledged AI assistant is the right tool for this? Specialized views could serve better? For example for your first question: we can put a "context view" for sessions which shows exactly what agent sees at that time? Everything we do via magic-context is deterministic so for each use case we should be able to find better UX than AI assistant. What do you think?

[zedd3v]

Wont this "context view" end up very large and unreadable? Purely due to amount of data/token?
And the context changes/prunes so often when youre working with large models/tasks that youd need some kind of state per message per session.

[ualtinok]

What I have in mind is to just show after message[0] which has the session history. So we'll only show current active context with tool calls collapsed (marking them via small badges if agent dropped those already, same for messages). It should be fairly readable if we do this. Context prunes are not often because those only happen when scheduler executes to keep cache as long as possible (around execute_threshold).