We study, theoretically and experimentally, the role of selective memory in shaping probability judgments.  In our model hypotheses and data act as cues that retrieve experiences from the memory database based on three regularities in human recall: frequency, similarity with the cue, and interference. We show that under different conditions on the similarity of experiences the model produces: overestimation of unlikely events, under-estimation of heterogeneous hypotheses, partition dependence, over-reaction, and under-reaction to data.  The model also produces belief heterogeneity from the same experiences and links it to heterogeneity in recall.  Experimental evidence confirms these predictions and finds a strong correlation between probabilistic beliefs and recall of exemplars.