Despite widespread acknowledgment of the importance of online semantic maintenance, there has been astonishingly little work that clearly establishes this construct. We review the extant work relevant to short-term retention of meaning and show that, although consistent with semantic working memory, most data can be accommodated in other ways. Using a new concurrent probe paradigm, we then report experiments that implicate a semantic maintenance capacity that is independent of phonological or visual maintenance that may build on a mechanism of direct semantic maintenance. Experiments 1 through 5 established that while subjects maintain the meaning of a word, a novel delay-period marker of semantic retention, the "semantic relatedness effect", is observed on a concurrent lexical decision task. The semantic relatedness effect refers to slowed response times when subjects make a lexical decision to a probe that is associatively related to the idea they are maintaining, compared to when the probe is unrelated. The semantic relatedness effect occurred for semantic but not for phonological or visual word-form maintenance, dissipated quickly after maintenance ends, and survived concurrent articulatory suppression. The effect disappeared when subjects performed our immediate memory task with a long-term memory strategy rather than with active maintenance. Experiment 6 demonstrated a parallel phonological relatedness effect that occurs for phonological but not semantic maintenance, establishing a full double dissociation between the effects of semantic and phonological maintenance. These findings support a distinct semantic maintenance capacity and provide a behavioral marker through which semantic working memory can be studied. (Contains 1 footnote and 9 figures.)