Because of the limited capacity of the short-term, working memory system texts must be processed one sentences or major clause at a time in cycles. This cyclical processing strategy eliminates the limited capacity problem but creates a new one; namely: How does a reader maintain the coherence of a text which must be processed in fragments? Kintsch and van Dijk (1978) have suggested that part of short-term memory is set aside as a buffer in which propositions from earlier processing cycles are held over the guide the integration of new propositions into the long-term representation of the text as a whole. Two experiments are presented which provide empirical support for this claim. Using the kintsch and van Dijk (Psychological Review, 1978, 85, 363-394) model to predict which propositions would be selected for inclusion in the short-term buffer, it was demonstrated that selected propositions are more available to subjects than unselected propositions from the same part of the text, and that the latter are no more availabe than unselected propositions from earlier in the text.