It is now well-established that long-term memory (LTM) knowledge, such as semantic knowledge, supports the temporary maintenance of verbal information in working memory (WM). This is for instance characterized by the recall advantage observed for semantically related (e.g. leaf - tree - branch) over unrelated (e.g. mouse - wall - sky) lists of items in immediate serial recall tasks. However, the exact mechanisms underlying this semantic contribution remain unknown. In this study, we demonstrate through a convergent approach involving computational and behavioral methods that semantic knowledge can be efficiently used to save attentional WM resources, thereby enhancing the maintenance of subsequent to-be-remembered items. These results have critical theoretical implications, and support models considering that WM relies on temporary activation within the LTM system.