Soft production preferences are a key feature of incremental language production, yet they lack a well-defined unified explanatory theory. Here, we propose an information-theoretic theory of availability effects grounded in the notion of lossy-context working memory, which takes the form of a cost function that can be applied to any computational-level model of language production. We show that production policies that minimize this cost function naturally give rise to key soft preferences observed in empirical data, including frequency bias, heavy-NP shift, and agreement attraction. We then show a novel prediction made by the model regarding the entropy of arguments' thematic roles, and show that this effect holds in corpus data.