The cognitive operations involved in the processing of surface-cohesion devices for the construction of a coherent mental representation is a major issue in text comprehension. An experiment was carried out with two narratives presented in two versions: a high-cohesion version and a low-cohesion version derived from the high version with the use of several devices--two anaphoric markers, changes in the temporal connective and word order, omission of the thematic sentence--without modification of the text content. The subjects read and immediately recalled a high-cohesion text and a low-cohesion text. The results showed that lowering cohesion produced an insignificant increase (8%) in reading time, but a highly significant decrease (25%) in recall performance. It appears that the subjects did not execute the processing required by the cohesion devices. The results are discussed with respect to models of sentence comprehension in comparison with text comprehension and metacognitive aspects of reading comprehension.