Orthogonal-Rotational Dynamics Supports Efficient Encoding and Updating for Streaming Information in Working Memory

We rely on the working memory (WM) to organize, store, and process the perpetual stream of information. Efficient encoding and processing of WM requires a framework that (1) separates individual memory items while accurately maintaining their temporal rank and (2) updates the sequence by discarding no-longer-needed items and accommodating newly arrived ones. To investigate the computational mechanisms underlying this functional implementation of WM, we analyzed the neural information representation in both a recurrent neural network (RNN) model and human subjects (<i>n</i> = 28, 18 males) under the same <i>N</i>-back WM task, which necessitates continuous encoding and updating of memory items. We discovered that an orthogonal-rotational dynamical framework facilitates memory encoding and updating, allowing both the RNN and brain to organize memory items efficiently. In the RNN model, we identified an orthogonal coding space where each memory item occupies a subspace corresponding to its ordinal rank. A rotational operation dynamically transfers information across these subspaces, updating memory while preserving their internal order. Overall, this orthogonal-rotational framework enables the network to store the information in a "first in, first out" manner. Remarkably, we also observed similar orthogonal-rotational dynamics in EEG signals recorded from the prefrontal areas of human participants engaged in the same task. These findings suggest a novel mechanism underlying the brain's ability to efficiently organize information stream for "online" processing and indicate that this strategy may be utilized by both biological and artificial neural networks for optimal information storage and updating.