In this work, we propose that working memory and time are "two sides of the same coin"—i.e., they are interwoven during encoding. To test this hypothesis, we designed two novel experimental paradigms by integrating temporal structures into classic working memory tasks (A).
For the first task (differential delayed-match-to-sample, dDMS), the duration of intervals was modulated based on the category of the initial input within a classic delayed-match-to-sample (DMS) framework. Although subjects were instructed to focus solely on working memory demands, the embedded temporal structure led them to subconsciously track time and anticipate the timing of subsequent inputs. Similarly, in the interval-stimulus-association task (ISA), subjects were required to respond based on the relationship between intervals and the second input, but they also implicitly retained the category of the first input.
To explore how temporal transformations influence neural encoding of working memory, we trained recurrent neural networks on three tasks derived from behavioral experiments (B):
We found that introducing temporal components shifted neural representations of working memory from stable attractors (persistent, constant firing) to dynamic attractors (temporally structured firing sequences). These dynamic attractors were shown to robustly encode working memory against external perturbations. Further analysis revealed that working memory and temporal information are co-encoded, computationally validating the hypothesis that they are "two sides of the same coin."
Related paper:
Zhou, S., Seay, M., Taxidis, J., Golshani, P., & Buonomano, D. V. (2023). Multiplexing working memory and time in the trajectories of neural networks. Nature Human Behaviour.
Address:No.138, medical college road 
Postcode:200032 
Tel/Fax:021-54237056
Emial:itbr@fudan.edu.cn