Abstract
The visual memory system encodes information selectively due to limited resources, resulting in systematic distortions and biases. Decades of previous work explains these biases in terms of categorical effects, whereby spatial memory is pulled towards perceptual attractors. Instead, we show that these biases can be accounted for by a rational process that retrieves information from representations with variable precision. Using a novel paradigm based on transmission chains, we explore these representations with an unprecedented degree of resolution. We show that internal memory representations are biased towards “semantic corners,” located near the boundaries between semantically meaningful regions within images. We show that these focal points are both diagnostic of image identity and associated with increased discrimination acuity. Our approach allows us to elucidate the nature of internal memory representations, providing a lens for understanding visual memory systems by revealing the latent internal geometry of their perceptual representations.