This research investigates how memory representations are activated and associated when making inferences in language, and in particular during the comprehension of anaphors (Le. co-referring expressions). Experiments 1 to 6 investigate 'do it' comprehension (e.g. John bought a newspaper. He did it while the others were out). Experiments 1 and 2 (offline sentence-completion tasks) show that 'do it' processing is sensitive to both NPs (a newspaper) and VPs (bought a newspaper) in the preceding context, and to specific lexical properties of the preceding NPs. With similar tasks, Experiments 3 and 4 show that the interpretation of an ambiguous 'do it' expression is influence by (two particular) properties of the linguistic context in which it is found. Experiment 5 (a reading-time study) suggests that 'do it' processing initially targets preceding NPs (and then only subsequently, preceding VPs), and Experiment 6 (an off-line grarnmaticality judgement task) shows that the 'do it' expression is a semantically divisible construction (i.e. 'do' + 'in, rather than an 'idiomatic' expression (Le. a semantically non-divisible 'do it'). In this way, Experiments 1 to 6 investigate how the referent of an anaphor is selected from the short term memory (STM) representation of a discourse. Experiments 7 to 11 however suggest that anaphor comprehension may also target the Mental Lexicon, a long term memory (LTM) store. From four on-line probe recognition tasks, (Experiments 8 to 11), and a lexical naming pre-test of the materials (Experiment 7), we find evidence to suggest that when an anaphor is processed, the meaning of the referent may be activated in some long-term linguistic storehouse of words (Le. similar in character to the Mental Lexicon).