Categorical storage of de-prioritized visual working memory content

Categorical storage of de-prioritized visual working memory content Surya Gayet, Dan Wang, Samson Chota, Stefan van der Stigchel Abstract Multiple items can be maintained in visual working memory (VWM) with different levels of priority (e.g., relevant now versus later). An emerging view, largely based on neuroimaging work, is that VWM content is stored in early visual cortex only when fine-grained detail is required. This occurs when items are prioritized, for instance to support precise recall or perceptual comparison. Instead, when VWM content is deprioritized, storage relies more on downstream regions, including parietal cortex where representations were shown to be more categorical. This framework leads to a clear prediction. If deprioritized items are stored in regions that do not support fine detail, then retrieving them into a prioritized state should reveal a loss of precision. Critically, such representations should also become more categorical. We tested this prediction using an existing dataset of three experiments (Wang et al., 2025). Participants memorized two colors for two consecutive delayed reproduction tasks using a continuous color wheel. One item was prioritized (cued for imminent report), whereas the other was temporarily deprioritized (relevant for later report). To quantify categorical structure, we conducted separate behavioral experiments (using the same experimental setup as Wang et al.) to identify color categories and map their prototypes and boundaries. Results showed that recall precision was lower for previously deprioritized items, indicating information loss. Also, color reports were generally biased toward (rather than away from) category prototypes, demonstrating influences of category membership on recall. Crucially, this bias was strongest for previously deprioritized items. These findings show that deprioritized VWM contents are stored in a more categorical format, consistent with a transfer from early visual cortex to higher-level regions. Although this flexibility supports efficient resource allocation, it comes at a cost: transfers are not lossless and systematically transform memory content. Categorical coding may support robust behavior in natural contexts, despite reduced precision.