Figure 2. Monkeys utilize context-dependent working memory by gating and overwriting representations within a shared subspace.

(A) Schematic illustration of three models for performing a context-dependent working memory task in low-dimensional subspaces. Representation of five stimulus locations (colored dots) used in the context of the Remember 1st and Remember 2nd task are shown in different PCA subspaces. The Pre-allocate model (middle left) suggests that representations are encoded in subspaces defined by task relevance, with task-relevant information consistently occupying the same subspace across different contexts. The Overwrite model (bottom left) proposes that stimulus representations may be either stored or overwritten in the same subspace, contingent on context. Diagrams on the right of each model illustrate the predicted decoding performance for each scenario. (B) Predicted decoding performance across time, across specific stimuli, and across Remember 1^st versus Remember 2^nd contexts for the three models. Each matrix element represents the decoding accuracy of a classifier trained at one time point and tested at another time point. (C, D) Empirical decoding performance of the first and second stimuli in (C) the prefrontal cortex and (D) the posterior parietal cortex. The first stimulus had two possible locations (chance = 0.5), whereas the second had five (chance = 0.2). For the cross-order and cross-context decoding analyses, we used the two shared locations across the two stimuli, yielding a chance level of 0.5. Cross-order decoding and cross-context decoding show that the information can be shared across order and task in the PFC. Solid lines show the empirical decoding accuracy, while dashed lines indicate chance level. Vertical lines separate a trial into five time epochs, as indicated in Figure 1: fixation, first stimulus, first delay, second stimulus, and second delay periods. (E) Cosine similarity between empirical decoding results and the predictions of the three models. The decoding profiles of the Overwrite models have higher cosine similarity with the PFC empirical data. The Overwrite model is also more similar to the PPC data except for cross-order decoding. (F) Context decoding performance (i.e., distinguishing Remember 1^st vs. Remember 2^nd contexts) in the prefrontal cortex and the posterior parietal cortex. The time point of zero refers to the onset of the first stimulus. Abbreviations R1: Remember 1^st; R2: Remember 2^nd; L1: the first stimulus location; L2: the second stimulus location.