In 4 experiments on retroactive interference (RI), we varied paired-associate learning lists that produced either appreciable or negligible forgetting. When the category of the stimulus word predicted its response word category, and the response was relatively unique within its category, learning was extremely rapid, and negative transfer and RI were negligible. The more the competing primed items in the predicted response category, the slower the learning and the greater the RI. If cues and responses were unrelated, learning was very slow, and RI was appreciable. Thus, predictive relations that help stimuli retrieve unique responses greatly alter forgetting in RI paradigms.