The purpose of this study was to investigate the learning mechanisms underlying motor adaptation of arm movements to externally applied perturbing forces. We considered two alternative hypotheses. According to one, adaptation occurs through the learning of a mapping between the states (positions and velocities) visited by the arm and the forces experienced at those states. The alternative hypothesis is that adaptation occurs through the memorization of the temporal sequence of forces experienced along specific trajectories. The first mechanism corresponds to developing a model of the dynamics of the environment, whereas the second is a form of "rote learning." Both types of learning would lead to the recovery of the unperturbed performance. We have tested these hypotheses by examining how adaptation is transferred across different types of movements. Our results indicate that 1) adaptation to an externally applied force field occurs with different classes of movements including but not limited to reaching movements and 2) adaptation generalizes across different movements that visit the same regions of the external field. These findings are not compatible with the hypothesis of rote learning. Instead, they are consistent with the hypothesis that adaptation to changes in movement dynamics is achieved by a module that learns to reproduce the structure of the environmental field as an association between visited states and experienced forces, independent of the kinematics of the movements made during adaptation.