Bedload movement is fundamentally probabilistic. Our quantitative understanding of gravel transport is particularly limited when flow conditions just exceed thresholds of motion, in part because of difficulties in measuring transport statistics during floods. We used accelerometer-embedded tracer clasts to precisely measure the timing of grain motions and rests during snowmelt floods in Halfmoon Creek, a gravel-bed mountain stream in Colorado, USA. These new data let us explore how probabilities of tracer movement vary with snowmelt discharge. Bedload hysteresis occurred over both daily and seasonal timescales, and included clockwise, counter-clockwise, and figure-eight patterns. We quantitatively explain these observations in terms of how thresholds of motion progressively evolved over 22 days during a seasonal snowmelt flood. Our results suggest that thresholds of motion are functions of both (a) cumulative shear stress and (b) temporal changes in shear stress during floods.