The context window is not your working memory
While building TokenBar, I noticed a weird behavior change in my own AI workflow.
As context windows got bigger, I got sloppier.
I stopped restarting chats.
I pasted in more raw tool output.
I carried old requirements and dead-end reasoning much longer than I should have.
Nothing looked broken, so it felt efficient.
It was not.
Three things got worse:
-
Review got slower
The model had more baggage, so I had more baggage to reread too. -
Prompt quality dropped
I relied on the model remembering old context instead of stating the current task cleanly. -
Spend got harder to predict
One long chat quietly turned into the default workflow.
What actually helped was not another after-the-fact dashboard.
It was seeing token usage live while I worked.
Once I could see the token trail in real time, I started using it as a workflow signal:
- restart instead of dragging stale context forward
- summarize logs instead of pasting everything back in
- stay on a smaller model until the task really needs more context
Big context windows are useful.
But they are easy to mistake for permission to be messy.
That lesson is a big reason I built TokenBar for macOS.
https://tokenbar.site/
