Atypical events are both memorable and unrepresentative of their class. We tested the hypotheses that (a) people tend to recall atypical instances of events, and (b) when they are unaware of this, they rely on these atypical instances in forecasting their affective reactions to future events. In three studies, participants who were asked to recall an instance of an event and participants who were asked to recall an atypical instance of an event recalled equally atypical instances. However, only the former participants made extreme forecasts about their reactions to future events. The results suggest that the impact bias (the tendency to overestimate the affective impact of future events) may be due in part to people's reliance on highly available but unrepresentative memories of the past.