How do different memory components impact risky choices? We developed a computational model that unifies compo-nents from memory research with decisions from experience. Our model chooses options based on expectations, observesoutcomes, stores them in memory, and forms new expectations based on observed outcomes. Their memory activationresults from recent encounters, binding outcomes to the context of options, and encoding according to similarity to exist-ing representations, and impacts how much each outcome drives new expectations. We tested the model on data from amulti-armed bandit task: Participants chose repeatedly between three options and received outcome feedback. Two coreoptions appeared in two choice sets with different third options. Core options were chosen less often when they wereaccompanied by similar (compared with dissimilar) third options. The model matched choice-proportion levels, direction,and size of this similarity effect. We present Bayesian estimates for memory components and discuss implications fortheory advancement.