How much text the model can hold at once — the whole conversation, your attachments and its own replies all live in here. When it overflows, the oldest turns get silently dropped.
0 used
Past half full, the oldest exchanges fold two-into-one on their own, in the background, while a running brief carries the thread. Turn the window down and it re-folds harder to fit, keeping the most recent 40% whole.
How long the model stays resident after your last message. Loading a 30B model costs several seconds, so keep it warm while you're working. Switching to a different model always unloads the one you were using, so only one is ever holding VRAM.
Tools execute on this machine. Everything but web access is confined to basalt/workspace.