Recent studies have suggested that during saccades cortical and subcortical representations of visual targets are represented and remapped in retinal coordinates. If this is correct, then the remapping processes must incorporate the noncommutativity of rotations. For example, our three-dimensional (3-D) simulations of the commutative vector-subtraction model of retinocentric remapping predicted centripetal errors in saccade trajectories between "remembered" eccentric targets, whereas our noncommutative model predicted accurate saccades. We tested between these two models in five head-fixed human subjects. Typically, a central fixation light appeared and two peripheral targets were flashed. With all targets extinguished, subjects were required to saccade to the remembered location of one of the peripheral targets and saccade between their remembered locations. Subjects showed minor misestimations of the spatial locations of targets, but failed to show the cumulative pattern of errors predicted by the commutative model. This experiment indicates that if targets are remapped in a retinal frame, then the remapping process also takes the noncommutativity of 3-D eye rotations into account. Unlike other noncommutative aspects of eye rotations that may have mechanical explanations, the noncommutative aspects of this process must be entirely internal.