LONG-LASTING WORKING MEMORY EMERGES FROM THE DYNAMICS OF HIERARCHICAL NEURAL ASSEMBLIES

Working memory is essential to cognitive functions and requires a neural signal capable of maintaining information over time. Classical models associate memory persistence with fixed firing rates in discrepancy with empirical observations from working memory experiments. Alternative frameworks propose that working memory is maintained through trajectories of population activity. However, these models often lack clear mechanisms for information readout, and long memory timescales require training. Here, we show that neural circuits with clustered connectivity, where neurons are organized into assemblies with strong intra-assembly connections, naturally generate long intrinsic timescales that support working memory without parameter tuning. We demonstrate that neural assemblies can persistently store a memory trace of transient inputs in their time-varying activity, which enables reliable decoding long after the input has ceased. This architecture allows neurons to remain dynamic and responsive, permitting persistent memory signals to coexist with time-varying activity and sequential inputs.
We examine how information propagates through hierarchical circuits with gradients of clustered connectivity. We find that memory can be transferred across such a chain, remaining decodable in downstream circuits long after it is no longer accessible in the upstream circuits receiving sensory inputs.
This architecture supports temporal filtering, robustness to noise, and sensitivity to new inputs, while preserving memory over long timescales. Because clustered assemblies are common in cortical organization, our results suggest that working memory is an inherent property of cortical circuits. These findings provide a circuit-level mechanism for working memory that reconciles persistence, dynamic activity, and efficient readout within a biologically plausible architecture.