Reviewer #2 (Public review):

Summary:

This work extends a previous recurrent neural network model of activity-silent working memory to account for well-established findings from psychology and neuroscience suggesting that working memory capacity constraints can be partially overcome when stimuli can be organized into chunks. This is accomplished via the introduction of specialized chunking clusters of neurons to the original model. When these chunking clusters are activated by a cue (such as a longer delay between stimuli), they rapidly suppress recently active stimulus clusters. This makes these stimulus clusters available for later retrieval via a synaptic augmentation mechanism, thereby expanding the network's overall effective capacity. Furthermore, these chunking clusters can be arranged in a hierarchical fashion, where chunking clusters are themselves chunked by higher-level chunking clusters, further expanding the network's overall effective capacity to a new "magic number", 2^{C-1} (where C is the basic capacity without chunking). In addition to illustrating the basic dynamics of the model with detailed simulations (Figures 1 and 2), the paper also utilizes qualitative predictions from the model to (re-)analyze data collected in previous experiments, including single-unit recordings from human medial temporal lobe as well as behavioral findings from a classic study of human memory.

Strengths:

The writing and figures are very clear, and the general topic is relevant to a broad interdisciplinary audience. The work is strongly theory-driven, but also makes some effort to engage with existing data from two empirical studies. The basic results showcasing how chunking can be achieved in an activity-silent working memory model via suppression and synaptic augmentation dynamics are interesting. Furthermore, we agree with the authors that the derivation of their new "magic number" is relatively general and could apply to other models, so those findings in particular may be of interest even to researchers using different modeling frameworks.

Weaknesses:

(1) Very important aspects of the model are assumed / hard-coded, raising the concern that it relies too much on an external controller, and that it would therefore be difficult to implement the same principles in a fully behaving model responsible for producing its own outputs from a sequence of stimuli (i.e., without a priori knowledge of the structure of incoming sequences).

(i) One such aspect is the use of external chunking cues provided to the model at critical times to activate the chunking clusters. The simulations reported in the paper were conducted in a setting where signals to chunk are conveniently indicated by longer delays between stimuli. In this case, it is not difficult to imagine how an external component could detect the presence of such a delay and activate a chunking cluster in response. However, in order for the model to be more broadly applicable to different memory tasks that elicit chunking-related phenomena, a more general-purpose detector would be required (see further comments below and alternative models).

(ii) Relatedly, and as the authors acknowledge in the discussion, the network relies on a pretty sophisticated external controller that decides when the individual chunking clusters are activated or deactivated during readout/retrieval. This seems especially complex in the hierarchical case. How might a network decide which chunking/meta-chunking clusters are activated/deactivated in which order? This was hard-coded in their simulations, but we imagine that it would be difficult to implement a general solution to this problem, especially in cases where there is ambiguity about which stimuli should be chunked, or where the structure of the incoming sequence is not known in advance.

(iii) One of the central mechanisms of the model is the rapid synaptic plasticity in the inhibitory connections responsible for binding chunking clusters to their corresponding stimulus clusters. This mechanism again appears to have been hard-coded in the main simulations. Although we appreciate that the authors worked on one possible way that this could be implemented (Methods section D, Supplementary Figure S2), in the end, their solution seems to rely on precisely fine-tuning the timing with which stimuli are presented - a factor that seems unlikely to matter very much in humans/animals. This stands in contrast with models of working memory that rely on persistent activity, which are more robust to changes in timing. Note that we do not discount the possibility of activity-silent WM, and indeed it should be studied in its own right, but it is then even more important to highlight which of its features are dependent on the time constants, etc.

(2) Another key shortcoming of this work is its limited direct engagement with empirical evidence and alternative computational accounts of chunking in WM. Although the efforts to re-analyze existing empirical results in light of the new predictions made by the model are commendable, in the end, we think they fall short of being convincing. As noted above, the model doesn't actually perform the same two tasks used in the human experiments, so direct quantitative comparisons between the model and human behavior or neural data are not possible. Instead, the authors rely on isolating two qualitative predictions of the model - the "dip" and "ramp" phenomena observed after a chunking cluster is activated (Figure 3), and the new magic number for effective capacity derived from the model in the case where stimuli are chunkable, which approximately converges with human recall performance in a memory study (Figure 4). Below, we highlight some specific issues related to these two sets of analyses, but the larger point is that if the model is making a commitment about how these neural mechanisms relate to behavioral phenomena, it would be important to test if the model can produce the behavioral patterns of data in experimental paradigms that have been extensively used to characterize those phenomena. For example, modern paradigms characterizing capacity limits have been more careful to isolate the contributions of WM per se (whereas the original magic number 7 is now thought to reflect a combination of episodic and working memory; see Cowan 2010). There are several existing models that more directly engage with this literature (e.g., Edin et al., 2009; Matthey et al., 2015; Nassar et al., 2018; Soni & Frank, 2025; Swan & Wyble, 2014; van den Berg et al., 2014; Wei et al., 2012), some of which also account for chunking-related phenomena (e.g., Wei et al, 2012; Nassar et al., 2018; Panichello et al., 2019; Soni & Frank, 2025). A number of related proposals suggest that WM capacity limits emerge from fundamentally different mechanisms than the one considered here - for example, content-related interference (Bays, 2014; Ma et al., 2014; Schurgin et al., 2020), or limitations in the number of content-independent pointers that can be deployed at a given time (Awh & Vogel, 2025), and/or the inherent difficulty of learning this binding problem (Soni & Frank, 2025). We think it would be worth discussing how these ideas could be considered complementary or alternatives to the ones presented here.

(i) Single unit recordings. We found it odd that the authors chose to focus on evidence from single-unit recordings in the medial temporal lobe from a study focused on episodic memory. It was unclear how exactly these data are supposed to relate to their proposal. Is the suggestion that a mechanism similar to the boundary neurons might be operative in the case of working memory over shorter timescales in WM-related areas such as the prefrontal cortex, or that their chunking mechanism may relate not only to working memory but also to episodic memory in the medial temporal lobe?

(ii) N-gram memory experiment. Our main complaint about the analysis of the behavioral data from the human memory study (Figure 4) is that the model clearly does not account for the main effect observed in that study - namely, the better recall observed for higher-order n-gram approximations to English. We acknowledge that this was perhaps not the main point of the analysis (which related more to the prediction about the absolute capacity limit M*), but it relates to a more general criticism that the model cannot account for chunking behavior associated with statistical learning or semantic similarity. Most of the examples used in the introduction and discussion are of this kind (e.g., expressions such as "Oh my God" or "Easier said than done", etc.). However, the chunking mechanism of the model should not have any preference for segmenting based on statistical regularities or semantic similarity - it should work just as well if statistical anomalies or semantic dissimilarity were used as external chunking cues. In our view, these kinds of effects are likely to relate to the brain's use of distributed representations that can capture semantic similarity and learn statistical regularities in the environment. Although these kinds of effects may be beyond the scope of this model, some effort could be made to highlight this in the discussion. But again, more generally, the paper would be more compelling if the model were challenged to simulate more modern experimental paradigms aimed at testing the nature of capacity limits in WM, or chunking, etc.

(iii) There are a number of other empirical phenomena that we're not sure the model can explain. In particular, one of the hallmarks of WM capacity limits is that it suffers from a recency bias, where people are more likely to remember the most recent items at the expense of items presented prior to that (Oberauer et al 2012). [There are also studies showing primacy effects in addition to recency effects, but the primacy effects are generally attributed to episodic rather than working memory - for example, introducing a distractor task abolishes the recency but not primacy effect]. But the current model seems to make the opposite prediction: when the stimuli exceed its base capacity, it appears to forget the most recent stimuli rather than the earliest ones (Figure 1d). This seems to result from the number of representations that can be reactivated within a cycle and thus seems inherent to the dynamics of the model, but the authors can clarify if, instead, it depends on the particular values of certain parameters. (In contrast, this recency effect is captured in other models with chunking capabilities based on attractive dynamics and/or gating mechanisms - eg Boboeva et al 2023; Soni & Frank (2025)). Relatedly, we're not sure if the model could account for the more recent finding that recall is specifically enhanced when chunks occur in early serial positions compared to later ones (Thalmann, Souza, Oberauer, 2019).

Figures and data

Illustration of the hierarchical working memory model.

Memory retrieval from a hierarchical structure.

Cognitive boundary neurons in the medial temporal lobe.

The new magic number bounds perfect-recall performance on verbal memory.

Simulation parameters in Fig. 1(d) and Fig. 2(b).

Full activity trace of the bottom panel in Fig. 1(d).

Additional RNN simulations with delayed Hebbian plasticity.

Individual 2D plots of Fig. 3(b).