Individuals of all ages and all cultures extract structure from the sequences of patterns they encounter in their environment, an ability that is at the very heart of cognition. One of the most widely accepted explanatory mechanisms that have been proposed is learning based on prediction. The idea is that individuals are constantly engaged in predicting upcoming patterns in their environment based on previously encountered patterns. Learning, in this view, is a process of gradually aligning these predictions with the outcomes that actually occur. Prediction-driven learning is the cornerstone of numerous computational models of sequence processing, and, in particular, the very well-known simple recurrent network (SRN, Elman, 1990). However, it turns out that prediction-driven models, in general, and the SRN, in particular, cannot account for a number of recent results in infant statistical learning and adult implicit learning. An alternative connectionist model, called TRACX (Truncated Recursive Autoassociative Chunk eXtractor), based, not on prediction, but on the recognition of previously (and frequently) encountered sub-sequences of patterns (chunks) will be presented that is able to handle empirical data that is problematic for prediction-based models. TRACX also accounts for a wide range of other empirical results. The main suggestion arising from this work is that recognition memory, not prediction, underlies sequence segmentation and chunk extraction.