Robots that need to count how many times they have done something often remember the past by recording frames over time. But time passes even when nothing happens. When a robot pauses or fidgets, its memory treats that idle time as progress, and it stops early.
Researchers tested this by making a robot briefly pause or move its wrist, actions that change nothing about the task. On two counting tasks, success dropped from 47 to 16 out of 50 trials on one task and a similar amount on another. The failure came from the memory, not the interruption itself: hiding the interruption from memory restored performance.
The robot was using elapsed time as a shortcut for progress, because its training demonstrations always ran at a steady pace. Removing time stamps did not help; the real problem was that memory was indexed by time. But indexing memory by events, recording only when something in the world changes, recovered most of the lost performance without retraining. This shows that memory meant to track what happened should advance with events, not with the clock.