Rhythmic replay of short-term memory neural patterns revealed by time-resolved error prediction

Theta oscillations are hypothesized to provide a temporal scaffold for short-term memory (STM). In this model, memory representations are organized into successive theta phases, reducing conflict between competing representations during encoding and maintenance. Previous studies have shown that sensory representations of memorized information are rhythmically reactivated at theta frequency. Whether this theta-rhythmic mechanism is shared across encoding and maintenance, traditionally viewed as temporally distinct processes, remains unclear. It is also unknown whether this rhythmic reactivation predicts subsequent memory fidelity rather than merely reflecting the reinstatement of sensory representations. We recorded EEG from 21 participants performing STM with colored, oriented items under low and high memory load. Using time-resolved multivariate pattern analysis, we predicted subsequent STM error from encoding- and maintenance-period activity. We found that distributed neural patterns predicted STM error and that this prediction fluctuated at theta frequency, independently of memory load and feature. Cross-temporal generalization indicated that a shared neural pattern recurred across encoding and into maintenance, consistent with periodic reactivation of a common neural code rather than a succession of distinct states. Prediction was temporally offset across spatial positions and item features, indicating that the features of a multi-feature object are encoded by separate, rhythmically interleaved processes. These findings characterize STM encoding and maintenance as a rhythmic, recurrent process and link theta-rhythmic fluctuations in memory fidelity to distributed cortical activity spanning multiple brain regions and frequency bands.