Attractor dynamics underlie sequence working memory in data-constrained recurrent neural networks

Sequences of stimuli can be maintained in working memory by encoding each item, together with its position, as a distinct pattern of neural population activity. Competing hypotheses about the underlying dynamics have been proposed: patterns of activity might be supported either by stable attractor states that persist over time, or by transient sequential dynamics. To investigate the dynamics underlying sequence memory, we developed a framework for fitting low-rank recurrent neural networks (RNNs) to multi-session recordings of spiking activity, and applied it to 2,488 units recorded across 30 sessions in two macaques. Our data-constrained RNNs generated population spiking activity which closely resembled neural data and recovered single-unit tuning properties across sessions. Consistent with prior work, fitted networks represented items presented at different sequence positions in separate low-dimensional subspaces in the neural activity, alongside a dominant temporal component. The RNNs enabled direct investigation of the underlying dynamics: By isolating the temporal and stimulus subspaces, we found a manifold of fixed points emerged during the memory retention period. Targeted perturbation analyses confirmed that activity within this manifold is consistent with attractor-like dynamics, and predict that brief optogenetic perturbations delivered along this attractive manifold can selectively induce persistent shifts in neural activity and decoded behavior. Together, our results indicate that sequence working memory is organized by position-specific coding subspaces, within which attractor states maintain persistent representations, while activity evolves along time-varying dimensions.