According to most theories of text comprehension, readers construct and store in memory at least two inter-related representations: a text base containing the explicit ideas in a text and a discourse model that contains the overall meaning or "gist" of a text. The authors propose a refinement of this view in which text representations are distinguished by both encoding and retrieval processes. Some encoding processes "unitize" concepts in a text and some "relate" units to one another. Units are retrieved based on familiarity processes in recognition, whereas related units are retrieved based on recollective processes. This distinction was tested in two experiments. In Experiment 1, readers comprehended sentence pairs in which some could be related by means of a causal inference, whereas others were only temporally related. Overall recognition was high in both conditions, but recollection, much more than familiarity, was sensitive to the causal manipulation. In Experiment 2, sentences began with a definite article as a linguistic cue to connect noun phrases or began with an indefinite article. The discourse manipulation had its primary influence on recollection. The authors suggest that the discourse model may be a collection of text ideas that are available to consciousness at retrieval. The gist-level representation of a text may not be a pre-stored structure; rather, it may be generated, in part, as a summary description of recollected text ideas. (Contains 2 figures, 3 tables, and 2 footnotes.)