Reports two experiments on mechanism of interference in working memory. Experiment 1 shows that a target word in a memory list, which bears high similarity to one of 4 words read aloud in the retention interval, is recalled less well than a control word. A second target word, not similar to any word read aloud but with all its phonemes repeated among the words read aloud, suffered the same degree of impairment. Feature overwriting explains both results, whereas similarity-based confusion explains only the first. In Experiment 2, participants remembered lists of 4 words, followed by 4 letters. A target word with high phoneme overlap with the letters was recalled worse than a control word, in line with predictions from feature overwriting. A further target word with many phonological neighbors created by inserting letter phonemes into it was recalled better than a control, contrary to predictions from feature migration.