Recent findings suggest a bidirectional relationship between preferences and choices such that what is chosen can become preferred. Yet, it is still commonly held that preferences for individual items are maintained, such as caching a separate value estimate for each experienced option. Instead, we propose that all possible choice options and preferences are represented in a shared, continuous, multidimensional space that supports generalization. Decision making is cast as a learning process that seeks to align choices and preferences to maintain coherency. We formalized an error-driven learning model that updates preferences to align with past choices, which makes repeating those and related choices more likely in the future. The model correctly predicts that making a free choice increases preferences along related attributes. For example, after choosing a political candidate based on trivial information (e.g., they like cats), voters' views on abortion, immigration, and trade subsequently shifted to match their chosen candidate.