We show how emergence offers new explanations for the behaviour of metaphorically-used expressions. Analysis of metaphors in two types of natural language data are combined: detailed analysis of continuous discourse, which offers wealth of context and the possibility of monitoring emergent forms as the discourse unfolds, and computer-assisted corpus analysis, which enables the examination of large numbers of examples of specific words and phrases across a range of contexts. We find that non-literal expressions with a relatively fixed form and highly specific semantics and pragmatics are very frequent in our data but are not well accounted for by current cognitive metaphor theory. We term these non-literal expressions ‘metaphoremes’, and argue that they represent the coalescence of linguistic, semantic, affective, and pragmatic forces into attractor states in the discourse system, appearing in discourse as relatively stable bundles of patterns of use. We show a metaphoreme emerging in the course of a discourse event and another which appears to have emerged recently as a result of a changing social environment. We then combine analyses and data types to track the use of <walk away from> as a metaphoreme, showing its patterns of formal, semantic, affective, and pragmatic characteristics.