Agent Harness Tutorial · End to end · Zoom level 2
This turn, in memory
While the loop runs, the harness holds the full working picture in memory. When the turn ends, almost all of it is thrown away on purpose.
Why throw the tool results away? Because they go stale. A claim that was "registered" ten minutes ago may be "under assessment" now. Saving it would tempt the model to answer from memory. Instead, rows read earlier are read again at the start of the next turn.
It also keeps the saved conversation small: four messages come to about 1.9 KB.