We used a prototype-distortion task and adopted ERPs to test between prototype and exemplar theories on categorization, which suggest that categories are represented as a prototype via an abstraction process or via storing previously encountered exemplars in memory, respectively. In Experiments 1 and 2, participants were presented low- or high-distortion category-members (i.e., dot-patterns) without anticipating subsequent categorization or recognition test. They were more likely to process low-distortions via prototype abstraction in both tests, and differently process high-distortions via prototype abstraction in categorization test, but via storing exemplars in recognition test. In Experiment 3, participants were explicitly instructed to do categorization or recognition task. We found that participants did categorization test via prototype abstraction (N1) only for studied items, not for unstudied items. And conversely did recognition via familiarity processing (FN400) only for category-members, not for non-members. In conclusion, the nature of category representations depends on the experimental contexts.