A prevalent assumption in text comprehension research is that many aspects of text processing are automatic, with automaticity typically defined in terms of properties (e.g., speed and effort). The present research advocates conceptualization of automaticity in terms of underlying mechanisms and evaluates two such accounts, a computational-efficiency account (underlying computational processes become more efficient with practice) and a memory-based processing account (the underlying basis of processing shifts with practice, from computing interpretations to retrieving prior interpretations). In five experiments, short texts containing either an ambiguous or unambiguous syntactic structure were presented for multiple study trials. In both conditions, reading times in target regions decreased across trials, indicating automatization. Several findings supported the memory-based processing account (e.g., practice effects were largely item-specific, reading times were longer for ambiguous versus unambiguous sentences on early trials but converged on later trials) Some evidence was also found for a contribution of gains in computational efficiency (i.e., some item-general practice effects were observed). Implications for research on automaticity and text processing are discussed.