In choice-reaction tasks, responses are faster if stimuli and responses are spatially compatible than if they are incompatible, even when the locations of the stimuli are irrelevant to the task. This stimulus-response (S-R) compatibility effect that occurs based on task-irrelevant stimulus and response features is known as the Simon effect. The Simon effect can be eliminated or even reversed after training with spatially incompatible S-R mappings only for a short duration, indicating that newly acquired incompatible S-R associations transfer to the Simon task. This transfer effect is usually reduced when the context of the training task is altered at test, suggesting that the expression of learned S-R associations depends on the similarity between the learning and test contexts. However, there can be cases where transfer occurs from one context to another but not in the reverse direction (i.e. transfer asymmetry). Transfer asymmetry is problematic for many models of psychological similarity, which would predict that transfer is symmetrical between two contexts. The present study shows that Tversky's set-theoretic model of similarity-the contrast model-is a useful framework for understanding how transfer symmetry arises in human perceptual-motor learning. The results of the two experiments imply that the similarity of contexts depends not only on features that overlap between the contexts but also on features that are distinctive to them.