It is now firmly established that long-term memory knowledge, such as semantic knowledge, supports the temporary maintenance of verbal information in Working Memory (WM). This support from semantic knowledge is well-explained by models assuming that verbal items are directly activated in long-term memory, and that this activation provides the representational basis for WM maintenance. However, the exact mechanisms underlying semantic influence on WM performance remain poorly understood. We manipulated the presence of between-item semantic relatedness in an immediate serial recall task, by mixing triplets composed of semantically related and unrelated items (e.g. leaf – tree – branch – wall – beer – dog; hand – father – truck – cloud – sky – rain). Compared to unrelated items, related items were better recalled, as had been classically observed. Critically, semantic relatedness also impacted WM maintenance in a complex manner, as observed by the presence of proactive benefit effects on subsequent unrelated items, and the absence of retroactive effects. The complexity of these interactions is well-captured by TBRS*-S, a decay-based computational architecture in which the activation occurring in long-term memory is described. The present study suggests that semantic knowledge can be used to free up WM resources that can be reallocated for maintenance purposes, and supports models postulating that long-term memory knowledge constrains WM maintenance processes.