Tell me what you read, and I will tell you what you remember: Evidence for personalized embeddings in memory modelling

Computational models of memory have achieved considerable success by formalizing how traces are encoded, cued, and retrieved. However, the role that individual linguistic experience plays in shaping representational structure has been relatively understudied. Where the issue has been examined, most models assume that a common population-level semantic space suffices, effectively treating individual variation in language experience as noise rather than signal. We test the theoretical adequacy of this assumption. In a large-scale experiment ( N = 478), participants completed a cued recall task in which identical target words were paired with cues drawn from eight different genre-specific semantic spaces (mystery, fantasy, horror, literary fiction, romance, science fiction, thriller, and non-fiction). Participants reported their reading habits across the eight genres, and personalized semantic representations were constructed by weighting genre-specific corpora proportionally to each participant's reading profile. Two complementary analyses were conducted. In a model-free analysis, cosine similarity computed within each participant's personalized semantic space predicted recall and omission outcomes more reliably than similarity derived from generic semantic spaces. In a computational analysis, substituting personalized representations into the embedded Computational Framework of Memory improved cue–target-level predictions over generic alternatives, accounting for approximately 4 percentage points more explained variance. These findings provide proof of concept that variation in individual linguistic experience shapes semantic structure in ways that are both behaviourally detectable and computationally tractable. More broadly, they suggest that the predictive limits of memory models may reside not only in their retrieval mechanisms but also in the fidelity of the representations they assume.