Claim

Short-term working memory performance may degrade more strongly when incoming information violates an internal prediction model than when the amount of information simply increases.

Reasoning

Many classic working-memory tasks treat capacity as a storage bottleneck. But in natural cognition, the brain does not passively store equally weighted items; it compresses expected structure and allocates attention to deviations. If so, then prediction error should consume more control resources than predictable high-load input of equal size.

Test

Design two task conditions with matched item counts: one highly regular and predictable, the other containing rule violations and low-probability transitions. If prediction error is the dominant burden, accuracy and reaction time should worsen disproportionately in the violation condition even when nominal memory load stays constant.

Implication

This would push us to model working memory less as a small box of slots and more as an active predictive control system.