The ability to actively remove irrelevant items from working memory storage is essential for goal-directed behavior. However, the mechanisms underlying active removal of information remain largely unknown. Importantly, exerting control over attentional selection of a memorized stimulus requires information about this specific selection. Our recent work provides evidence that this information is represented in the form of selection-specific activity in parietal and temporal regions. Here, we employ fMRI and MVPA to identify selection-specific activity across the cortex through a multimodal retro-cued n-Back discrimination task. Subjects were tasked with memorizing an item, either visual or auditory, until a colored retro-cue indicated if that item should be remembered or forgotten. Thus, while prior work investigated which process underly the selection between two items, the current study focuses on the fate of a singular item. Selection suppressed the representation of the to- be-forgotten item in visual and parietal cortex, as indicated by reduced item-specific decodability. To identify areas that carry selection-specific information, whether an item should be actively removed from working memory, we performed a whole-brain searchlight analysis (cvMANOVA for ‘remember’ versus ‘forget’) for the period following cue onset. We find a wide array of frontal and parietal brain regions holding this information, including precentral and lateral prefrontal cortex as well posterior and medial parietal regions, across both stimulus modalities. These results highlight that active removal of working memory contents occurs in both sensory and non-sensory cortical regions and reveal a generalized, modality-independent mechanism for cognitive control of information stored in working memory.

Working memory (WM) prioritization supports flexible, goal-directed behavior by selecting task-relevant information among competing representations. Bursts of neural oscillations have been proposed to contribute to WM control (Lundqvist et al., 2018), but their specific role in prioritization remains unclear.

To address this, we recorded local field potentials from prefrontal cortex (PFC) and visual area V4 in two non-human primates performing WM selection and spatial attention tasks. In both tasks, subjects viewed two colored squares. In the selection task, a spatial cue indicated whether the color at the upper or lower location had to be reported after a post-cue memory delay (~600 ms). Spatial attention task had the same structure but the cue preceded the stimulus presentation thus directing attention to visual input.

High-frequency oscillatory bursts reliably distinguished whether the upper or lower item was prioritized in WM. These patterns generalized across WM and attention tasks, consistent with domain-general prioritization mechanisms. Furthermore, control-related spatial signals reshaped how the selected color was represented in WM, changing encoded color representations from location-specific before the cue to location-independent post-cue.

Interestingly, cue-encoding channels formed ensembles that transiently synchronized at low frequencies across PFC and V4, and this synchronization modulated the reformatting of WM representations during prioritization. Together, these results show that high-frequency bursts across PFC and V4 track the dynamic reformatting of WM representations during selection. Such flexible, item-specific prioritization is supported by transient temporal alignment of oscillatory activity across areas.

Interacting with moving objects in our environment, such as while driving or playing sports, requires anticipating their future locations across space and time, especially when they become temporarily occluded from view. How do we predict the location of a moving object that is no longer visible? Existing accounts propose that information encoded during visible motion is used to extrapolate trajectories in the absence of corresponding visual input. While prior work shows that early visual cortex represents both past and anticipated events, it remains unclear whether it maintains a continuous spatial representation of object location during occlusion. Here, we used fMRI to test whether visual cortex supports sustained spatial updating during occlusion and whether such representations are modulated by attention. Participants viewed a target moving along a circular trajectory while performing three tasks: fully visible motion, partially occluded motion requiring judgments about object’s location, and an attention-at-fixation task during identical partially occluded stimulus. We found that visual cortex continuously updated representations of the occluded object’s location for up to four seconds. Diverting attention away from the moving stimulus attenuated, but did not eliminate, these representations. Moreover, the strength of these representations in early visual cortex was associated with behavioral uncertainty in estimating object location during occlusion. Together, these findings support a tracking-based account of motion extrapolation and show that visual cortex encodes continuously updated spatial information about occluded objects.

Working memory has traditionally been characterized as a system for the short-term storage of sensory information. However, its functional purpose extends beyond reflecting the past: working memory serves as a core cognitive function that bridges potentially relevant visual input and anticipated future actions. It is therefore crucial to understand not only how sensory information is encoded and retained in working memory, but also how it is prepared for future use. Across three complementary EEG studies, we investigated how visual representations and their associated action plans are integrated into working memory, and how this integrated code is maintained and reinstated following external interference. We developed a visual-motor task that, combined with time-frequency analyses, allowed us to independently track visual and motor representations via modulations of posterior alpha (8-12 Hz) activity relative to the memorized item location and central beta (13-30 Hz) activity relative to the required response hand. We show that prospective action plans do not emerge gradually during memory delays, but are incorporated into working memory early, in tandem with visual encoding. This early action encoding precedes a second stage of action preparation that adapts to the expected timing of memory utilization and occurs even in advance of an intervening external task. Following external interference, visual and motor attributes are rapidly and concurrently reselected after completion of the external task, rather than only just in time for memory-guided behavior. Together, these findings demonstrate that working memory stores a dual visual-motor code that supports robust, action-ready representation optimized for future use.