Flexible Working Memory in the Peripheral Nervous System

Working memory (WM) representations that are distributed across the brain can be flexibly recruited to best guide behavior ^1–4 . For instance, information may be represented relatively more strongly in visual cortex when a WM task requires fine visual detail, or more strongly in motor cortex when a specific response can be prepared ^5–10 . If WM is geared to prospectively guide actions, then we might also expect such task-oriented neural signals to propagate to the peripheral sensors and effectors that realize WM goals. Likewise, there is now evidence that oculomotor signals like saccade biases can track simple visuo-spatial WM features ^11,12 . However, it is unclear if such signatures are functionally meaningful, and how much information is contained in them. Here, we test the idea that WM content is adaptively distributed across the nervous system according to behavioral demands. Namely, we test whether visual WM stimulus features are expressed in patterns of both eye and hand movements during a WM delay, and whether the distribution of such peripheral motor activity shifts with the task context. In a delayed recall task, we manipulated how human participants reported their memory by having them either draw a line or adjust a wheel to match a remembered orientation. Via eye- and stylus-tracking, we found that remembered orientations were decodable from small inflections in both gaze and hand movements during the blank WM delay. Moreover, this decoding strength varied by response format: gaze patterns tracked memorized features relatively better in the wheel condition (vs. draw ), while hand movements were better in the draw condition (vs. wheel ). Individuals who showed greater wheel benefits in gaze-based decoding also showed greater draw benefits in hand-based decoding, suggesting a strategic processing shift to the more relevant system. Therefore, visually encoded WM contents may be adaptively allocated to the most task-relevant motor effectors, balancing WM representations across peripheral activity according to behavioral needs.