Working memory templates are maintained as feature-specific perceptual codes.

Working memory (WM) representations serve as templates that guide behavior, but the neural basis of these templates remains elusive. We tested the hypothesis that WM templates are maintained by biasing activity in sensoriperceptual neurons that code for features of items being held in memory. Neural activity was recorded using event-related potentials (ERPs) as participants viewed a series of faces and responded when a face matched a target face held in WM. Our prediction was that if activity in neurons coding for the features of the target is preferentially weighted during maintenance of the target, then ERP activity evoked by a nontarget probe face should be commensurate with the visual similarity between target and probe. Visual similarity was operationalized as the degree of overlap in visual features between target and probe. A face-sensitive ERP response was modulated by target-probe similarity. Amplitude was largest for probes that were similar to the target, and decreased monotonically as a function of decreasing target-probe similarity. These results indicate that neural activity is weighted in favor of visual features that comprise an actively held memory representation. As such, our findings support the notion that WM templates rely on neural populations involved in forming percepts of memory items.