The role of semantic codes in verbal working memory maintenance: an fMRI investigation

Although the impact of long-term memory (LTM) knowledge on working memory (WM) performance has been well demonstrated, the nature of the interactions between WM and semantic LTM remains poorly understood. While some theoretical models consider that this intervention indicates a temporary and direct activation of LTM during WM maintenance, others consider that LTM only intervenes during recall, in order to reconstruct degraded memory traces. The aim of this study was to examine the extent to which semantic knowledge in LTM is temporarily activated during the maintenance phase in WM, by determining, via fMRI, whether brain markers of semantic knowledge are observable during the maintenance phase in verbal WM. Twenty-seven healthy participants (18-35 years old) participated in this study. The multivariate brain signals of four semantic categories were pre-determined via an implicit semantic activation task (reading aloud of words from four semantic categories). Next, the participants were asked to maintain words in a verbal WM task, these words being the names of the four semantic categories implicitly activated in the preceding task. Via multi-voxel pattern analyses, we were able to distinguish the neural patterns associated with the four semantic categories both in the semantic activation task and during the maintenance stage of the WM task (Bayesian factors BF10>1,078 e+7 and BF10>2.058, respectively, for category decoding against chance level decoding). However, when attempting to predict the neural patterns associated with the semantic categories between the two tasks, prediction rates did not exceed chance level, indicating that the multivariate brain signals associated with the maintenance of semantic category names in the WM task and those associated with semantic categories in the implicit semantic activation task differ. These results indicate that deep semantic representations in LTM are not necessarily activated during maintenance in WM and suggest a more limited intervention of semantic LTM in WM.