In the current research, we tested the idea that the proximity of contrasting categories in a learning sequence would determine the features to which participants attend in a categorization task. For the first experiment, we designed a 4-category structure in which pairs of categories could be perfectly distinguished using 1 feature. Two of the categories were paired together in the first part of the learning phase, followed by the other 2 categories in the second part of this phase. In a transfer test in which all 4 categories were shown, participants attended more to the features that differentiated the paired categories than to the other, equally diagnostic features. In the second experiment, we extended this finding to a task that involved all 4 categories but in which pairs of categories were more likely to be interleaved. Once again, participants were more likely to pay attention to the dimensions that separated the 2 categories in proximity in the sequence. These findings suggest that the local learning context influences the representation of a category.