Probabilistic working memory representations in human cortex guide behavior

Models of working memory make fundamentally different commitments to the architecture of individual memories. Information-sparse models conceptualize individual memories as single point estimates agnostic to meta-cognitive variables such as uncertainty. In contrast, information-rich models propose memories are encoded as probability distributions over feature space that embed memory uncertainty in the shape of the distribution. To distinguish these accounts, we constructed probability distributions of memory from participant's iterative reports of motion direction on each trial. Remarkably, the idiosyncratic shape of these distributions (e.g., asymmetry) on single trials matched the shape of neural probability distributions decoded from fMRI patterns measured from occipital and parietal cortex. Consistent with information-rich models, the neural representation of an individual memory encodes more than the memorized feature; its variance (i.e., width and asymmetry) encodes idiosyncrasies whose read-out predicts memory behavior.