When a situation could lead to multiple mutually exclusive consequences, recent research shows that people automatically generate multiple predictive inferences in memory. Several theoretical mechanisms have been proposed to account for the generation of predictive inferences. One hypothesis is that inferences are minimally encoded, represented only by a set of semantic features related to the inferences or by a more general concept that covers the consequences of both inferences. A second hypothesis is that activation of the inferences is delayed because it is initially weak and requires time to build. In three experiments designed to examine these two hypotheses, participants read narrative passages that supported mutually exclusive consequences. Predictive inferences were not detected in working memory using a word-naming task, even at longer delays, but were detected using a short-term contradiction paradigm and, therefore, available as part of the discourse representation. The combined results indicate that predictive inferences are minimally encoded rather than their activation being delayed.