Category representations are often assumed to reflect the statistical distribution individual category members. However, recent work shows that people’s category representations tend to be biased toward high-value category members. We propose that this bias stems from prioritized memory: when learning about a category, people devote more cognitive resources to remembering important or desirable items, leading to their overrepresentation in category-level representations. We test key predictions of this account behaviorally and computationally. Behaviorally, we find a strong correlation between the features people prioritize in memory and the features that dominate their spontaneous recall of category members. Computationally, we use variational autoencoders to show that when statistical learning prioritizes accuracy for certain items, these items are overrepresented when sampling from the learned category distribution. Together, these findings suggest that prioritized memory plays a key role in shaping category representations.