Loading slide
You just watched it happen on the board. Now let us name what it means.
When a conversation runs long enough to fill the window, something has to leave. The tool builders get to choose how. Most often the rule is the simplest one: wipe the oldest turns first, making room at the bottom by erasing from the top. Some tools instead pause and write a short summary of the early conversation, then keep the summary and drop the originals. A few just refuse to go on until you start fresh.
Whichever rule is used, the effect on the model is the same. It can only reason about what is still on the board. Push the part where you said "keep it brief" off the top, and the model starts writing long again, not out of stubbornness, but because that instruction genuinely is not in front of it anymore. It might repeat a suggestion it already made. It might break a rule you set twenty messages ago. From its side, none of that ever happened.
This is not a bug to be patched. It is the shape of the thing. The board is finite, so a long enough conversation will always spill over its edge.
That points to one practical habit, and it is worth holding onto. Anything you need the model to obey all the way through a long task has to stay on the board the whole time. The safest place is the standing instructions at the top, which do not get wiped. The next safest is your own habit of restating the key constraint every so often, so a fresh copy keeps landing back on the board before the old one slides off.