Abstract
Are working memory representations point estimates (e.g. a specific shade of red) or are they richer, consisting of a probability distribution over possible features? In a color memory task, subjects reported on each trial a segment of the color wheel that reflected both their estimate of the memorized color and their uncertainty. The reported uncertainty correlated with estimation error. In blocks in which a non-uniform prior over colors was used, estimates were biased towards the prior mean, and the bias was larger for larger reported uncertainty. Our findings are consistent with a Bayesian framework in which people reason with probabilistic color memories.