Learning a spoken language requires learning a phonological inventory and phonotactics, or the sequences of phonemes possible in the language. Laboratory investigations of phonotactic learning include tongue-twister studies that show that speech errors respect artificial phonotactic constraints, for example that /k/ never appears as a syllable onset. The current research investigates whether errors can reveal similar learning in nonlinguistic domains, specifically in immediate serial recall studies. In Experiments 1-3 participants recalled sequences of 6 items, grouped into two 3-item subsequences, in verbal immediate serial recall experiments. Some items were restricted to appear in specific sequence positions while others were unrestricted, as a parallel to the artificial phonotactics learning experiments. As with speech errors, recall errors with restricted items showed sensitivity to the statistical regularities built into the experiment. Similar results are shown for both absolute and probabilistic constraints. However, learning effects were weaker than in the phonological domain, and were less influenced by hierarchical structure than what is observed in phonotactic learning (Experiment 4). Taken together, these results suggest that there are both domain-general and language-specific constraints on phonotactic learning.