Woohyeuk (Leo) Chang¹, Ed Awh¹; ¹University of Chicago

Classic models propose that information retrieved from long-term memory (LTM) is reinstated into working memory (WM), engaging the same capacity-limited operations that support online storage. For example, alpha oscillations decline as a function of the number of items stored in WM, and a similar pattern is observed when observers retrieve multi-item arrays from LTM. Although this suggests that LTM retrieval mirrors WM encoding, other work indicates that alpha activity may instead reflect the spatial extent of covert attention rather than the number of WM representations per se. Thus, there is strong motivation to study reinstatement using neural signals more directly linked to WM storage and designs that de-confound spatial attention from the number of stored items. We used a sequential design in which items were separated in time rather than space, holding spatial attention constant while manipulating the number of individuated items. A multivariate decoding approach known to track WM load independently of spatial attention was used. Participants memorized LTM shape arrays containing 1 or 3 targets and then completed a sequential WM/LTM task during EEG recording. Voltage-based signatures of WM load generalized across conditions requiring the maintenance of temporal order and item identity. The key question was whether retrieving multi-item arrays from LTM would elicit a multi-item WM load signature or whether prior learning would result in unitization (i.e., chunking). Consistent with the latter, LTM retrieval elicited the same neural pattern as a WM load-1 trial, regardless of the array’s size. Alpha power was likewise insensitive to the number of retrieved items, likely because each sequence appeared at a single location. Thus, retrieval of multi-item arrays from LTM reveals a unitized representation that is insensitive to set size, providing insight into how associative learning shapes the reinstatement of multi-item LTM structures in WM.