Summary: Sensory evidence tends to be fleeting, often unavailable when we categorize or estimate world features. To overcome this, our brains sustain sensory information in working memory. Although keeping that information accurate while acting on it is vital, humans display two canonical biases: estimates are biased toward a few stimuli (“stimulus-specific bias”) and prior decisions (“decision-consistent bias”). Integrative—especially neural mechanistic—accounts of these biases remain scarce. Here, we identify drift dynamics toward discrete attractors as a common source of both biases in orientation estimation, with decisions further steering memory states. Behavior and neuroimaging data reveal how these biases co-evolve through the decision-steered attractor dynamics. Task-optimized recurrent neural networks suggest neural mechanisms that enable categorical decisions to emerge from working memory for continuous stimuli while updating its trajectory, warping decision-consistent biases under stimulus-specific drift.
How early visual cortex (EVC) represents working memory (WM) content while continuing to process incoming sensory input remains unclear. Using fMRI with prolonged delays to isolate mnemonic activity, together with analyses of cross-decoding, low-dimensional subspace structure, and representational geometry, we examined the relationship between sensory and mnemonic orientation representations in human EVC. Cross-decoding generalized poorly between sensory and mnemonic epochs, but this did not imply unrelated codes. Rather, the two occupied separable low-dimensional subspaces while preserving representational geometry across epochs. During discrimination and estimation, sensory- and mnemonic-trained decoders yielded dissociable readouts of concurrent sensory and mnemonic information from the same EVC measurements. Mnemonic coding showed little dependence on the retinotopic radial bias that characterized sensory coding, and trial-by-trial variability in mnemonic representation predicted both discrimination choices and estimation reports. Our findings support a population-level account in which mnemonic information in EVC is re-expressed in a separable but geometrically preserved format.