No laboratory test can reliably identify patients with schizophrenia. Instead, key symptoms are observed via language, including disorganized language and delusions. Underlying brain processes remain unclear; characterizing them would greatly enhance our understanding of schizophrenia. In this situation, computational models can be valuable tools to formulate testable hypotheses. This work aims to capture the link between biology and symptoms using DISCERN, a connectionist model of storytelling. Candidate illness mechanisms for schizophrenia are simulated and evaluated through narrative language, i.e. at the same level used to diagnose patients. The result is the first simulation of abnormal storytelling in schizophrenia, both in acute psychotic and compensated stages. Of all illness models, hyperlearning, a model of over-intense memory consolidation, produced the best fit to language abnormalities of stable outpatients, as well as compelling models of psychotic symptoms. If validated experimentally, the hyperlearning hypothesis could provide a platform for developing future treatments for schizophrenia.