Corpus studies by Schuler, AbdelRahman, Miller, and Schwartz (2010), appear to support a model of comprehension taking place in a general-purpose working memory store, by providing an existence proof that a simple probabilistic sequence model over stores of up to four syntacticallycontiguous memory elements has the capacity to reconstruct phrase structure trees for over 99.9% of the sentences in the Penn Treebank Wall Street Journal corpus (Marcus, Santorini, & Marcinkiewicz, 1993), in line with capacity estimates for general-purpose working memory, e.g. by Cowan (2001).But capacity predictions of this simple structure-based model ignore non-structural dependencies, such as long-distance fillergap dependencies, that may place additional demands on working memory.Distinguishing unattached gap fillers from open attachment sites in syntactically-contiguous memory elements requires this contiguity constraint to be strengthened to a constraint that working memory elements be semantically contiguous.This paper presents corpus results showing that this stricter semantic contiguity constraint still predicts working memory requirements in line with capacity estimates such as that of Cowan (2001).