Human verbal list memory is inherently imperfect, often resulting in extralist errors-the incorrect recollection of unstudied information. While existing serial recall models have their strengths, they struggle to account for these imperfections due to the absence of a crucial component: a lexicon with structured representations. We address this limitation with the Embedded Computational Framework of Memory, an approach that integrates a lexicon into a recall model. This integration captures intrinsic similarities between studied and unstudied information, enabling word-specific predictions that reflect participants' behavior. We conducted computational demonstrations using large single-trial experiments (350 and 550 participants) and a multitrial serial recall experiment (1,000 participants). These demonstrations involved lists that were semantically (Demonstrations 1, 2, and 5), phonologically (Demonstration 3), or orthographically (Demonstration 4) related or unrelated, to assess the distribution of extralist errors. We used several variants of Embedded Computational Framework of Memory, each incorporating lexicons that captured either semantic, phonological, or orthographic information-or, for comparison, a random lexicon. The Embedded Computational Framework of Memory accounts for typical serial recall performance-including correct recall, intralist errors, and omissions-while also capturing the pattern of extralist errors. Furthermore, it uniquely evaluates memory errors at both the list and item levels, and delineates the relationship between the number of related words/nonwords and extralist errors across orthographic, phonological, and semantic information. This solution is compatible with most memory models and supports a shift from arbitrary to structured word representations in computational modeling, as a necessary step toward understanding why people recall unstudied words. (PsycInfo Database Record (c) 2026 APA, all rights reserved).
Guitard et al. (Mon,) studied this question.