We offer a theory of decisions in which attention to the features of choice options is determined by the DM’s categorization of the current choice problem in a set of problems she solved in the past. Categorization depends on goal-relevant as well as contextual problem-level features. The model yields systematic heterogeneity in attention and choice in a given problem based on dierent past experiences, rigidity of choices when categorization does not change despite new data, and discontinuous shifts when changes in bottom-up salient features cause re-categorization. The model unies major puzzles and framing effects in riskless, statistical, and lottery choice based on heterogenous and unstable mental representations.