This paper proposes a model of choice under risk based on imperfect memory and self-deception. The model assumes that people have preferences over their own attributes and can, to some extent, manipulate their memories. It leads to a non-expected utility representation and provides a uni ed explanation for several empirical regularities: non-linear probability weights, small-stakes risk aversion, regret, and the competence hypothesis. It also leads to endowment and sunk cost e¤ects. The model implies that behavior will converge to the one predicted by expected utility theory after a choice has been made a su¢ ciently large number of times.