In recent years, it has come to be widely accepted that linguistic and situational context must be used in understanding an utterance. It is not at all clear, however, how to represent such contextual information or what its precise role in the understanding process is. In fact, even the very nature of context remains problematic. In this paper, we look at task-oriented dialogue² and propose some initial answers to these questions. Our answers take the form of several distinctions and structures, from which we draw some broad implications. The reader is cautioned that we intend our observations to be suggestive rather than in any way exhaustive. (See, for example, Grosz [4] for a different, but complementary, perspective on this problem.)