We studied how subjects learned to make movements against unpredictable perturbations. Twelve healthy human subjects made goal-directed reaching movements in the horizontal plane while holding the handle of a two-joint robotic manipulator. The robot generated viscous force fields that perturbed the limb perpendicular to the desired direction of movement. The amplitude (but not the direction) of the viscous field varied randomly from trial to trial. Systems identification techniques were employed to characterize how subjects adapted to these random perturbations. Subject performance was quantified primarily using the peak deviation from a straight-line hand path. Subjects adapted their arm movements to the sequence of random force-field amplitudes. This adaptive response compensated for the approximate mean from the random sequence of perturbations and did not depend on the statistical distribution of that sequence. Subjects did not adapt by directly counteracting the mean field strength itself on each trial but rather by using information about perturbations and movement errors from a limited number of previous trials to adjust motor commands on subsequent trials. This strategy permitted subjects to achieve near-optimal performance (defined as minimizing movement errors in a least-squares sense) while maintaining computational efficiency. A simple model using information about movement errors and perturbation amplitudes from a single previous trial predicted subject performance in stochastic environments with a high degree of fidelity and further predicted key performance features observed in nonstochastic environments. This suggests that the neural structures modified during motor adaptation require only short-term memory. Explicit representations regarding movements made more than a few trials in the past are not used in generating optimal motor responses on any given trial.