When we practice a motor task, we can do it better the next time we revisit it. How is this accomplished? The basic assumption in neuroscience has been that during practice, we learn an association between the stimulus and the appropriate motor commands. However, it has been difficult to reconcile this view with two basic behavioral results: (1) when learning is followed by a long period of washout (removal of the perturbation), the motor memory appears protected from erasure (termed “savings”). How is it that learning followed by washout does not erase the association between stimulus and motor commands? (2) When washout is following by learning of the opposite perturbation, subjects exhibit meta-learning, i.e., performance is better than naïve in a perturbation opposite to the one that they had initially learned. How could learning to associate a stimulus to one direction of motor commands followed by washout help in learning in the opposite direction? Here, we approach these problems from a new perspective: the neural architecture that supports motor learning in the cerebellum. We propose that in the cerebellum, micro-clusters of Purkinje cells (P-cells) are organized based on their preference for error. This preference is expressed in their complex-spike tuning (encoding of error), which in turn provides a coordinate system in which simple spikes can be understood. The P-cell’s error preference makes it so that when error changes, anatomically distinct P-cell micro-clusters are recruited. As a result, when a perturbation is followed by washout, error changes direction and engages new groups of P-cells, producing a new memory without erasing the old. The same hypothesized anatomy suggests that meta-learning arises not because of similarity of the motor commands, but because of the similarity of errors. Using this hypothesis we show that when simple spikes of P-cells are organized into micro-clusters, an exquisite encoding of motion emerges. We propose to test a host of predictions regarding both the neurophysiological correlates of error-dependent learning in the cerebellum, and its behavioral correlates of savings and meta-learning in healthy people. Finally, we use the theory to better understand a latent form of motor learning in people with damage to their cerebellum. From a clinical perspective, our work aims to understand how the brain stores motor memories, and how it up-regulates learning from error, questions that are relevant to motor rehabilitation following neuro- trauma and disease. Our theory provides a recipe to modulate error-sensitivity, which should produce faster motor learning, potentially affecting the duration of rehabilitation.