Five experiments, using a probe recognition procedure, examined the codes involved in short-term memory. Interference effects were obtained for probes with an orthographic, associative, or semantic relationship to a memory set item. The orthographic interference effects appeared to result from matches on the basis of abstract letter codes (i.e., graphemes) rather than overlap of visual features. The associative and semantic effects differed from what might be predicted on the basis of spreading activation, suggesting that these effects are qualitatively different from semantic priming effects observed in lexical decision tasks. These results support models of short-term memory in which items are represented as groups of features and recognition depends on a comparison of the features of the probe to those of the memory set items.