Working memory (WM) and its capacity have been extensively studied in controlled tasks, where information is briefly presented and followed by a memory probe. In contrast, real-world environments often allow sensory information to remain accessible, letting individuals choose between relying on external sensory inputs or using memory representations. Interestingly, WM usage in such natural settings is lower than expected from traditional capacity estimates. Further, existing studies reveal considerable variance in how much participants engage WM naturally, but little is known about whether this reflects stable individual differences or what factors influence natural WM usage. Here, using two independent datasets from studies utilising an object copying task in VR, we demonstrate that individual differences in WM usage are stable across time and task context. Next, in a preregistered study (n=100), we investigate two potential predictors of natural WM use: WM capacity and attentional control. Participants completed an VR object copying task to assess WM usage, and three WM span and attentional control tasks. Both WM capacity and attentional control independently predicted natural WM usage. Our approach represents the first systematic exploration of factors determining natural memory reliance in complex behaviour, showing how benchmarks of cognitive function relate to everyday cognitive engagement.