A working memory model integrating meaning

Verbal working memory (WM) is strongly influenced by semantic similarity between items. Serial recall of lists of semantically similar words is better than of dissimilar words. Despite being a well-replicated phenomenon, this support of semantic knowledge to WM hasn’t been modeled in an exhaustive manner. We present a connectionist WM architecture integrating meaning. The core WM architecture encodes distributed representations of items and contexts through temporary binding. The items are assumed to be represented in a purely phonological and/or orthographic format. An important assumption of the model is that items’ meaning is not directly encoded into WM through item-context binding. Instead, semantic knowledge supports WM through sustained activation in a long-term memory semantic network that we supplemented to the core WM architecture. This way of integrating meaning into WM accounts for the impact of semantic similarity on WM performance in different experimental manipulations, including: (1) the beneficial effect of semantic similarity on item memory, (2) the absence of impact of semantic similarity on order memory, (3) the protective effect of semantic similarity against interference and (4) the reduced semantic similarity effect as the positional distance between similar items increases. Based on these results, we argue that meaning doesn’t have to be directly encoded into WM. Instead, we propose an alternative view according to which semantic knowledge supports WM through activated long-term memory.