Skip to main content
NCBI home page
Human Brain Mapping logoLink to Human Brain Mapping
. 2026 Jun 30;47(10):e70597. doi: 10.1002/hbm.70597

Intrinsic Functional Architecture Reflects Individual Differences in Passive Working Memory: An Exploratory Resting‐State fMRI Study

Yun Tian 1, Jiangtao Chen 2, Ziyuan Li 3, Li Gong 3, Qiang Liu 1,✉
PMCID: PMC13316451  PMID: 42376739

ABSTRACT

Passive working memory (WM) is rarely detectable because it is thought to rely less on persistent neural firing, leaving a minimal trace in ongoing brain activity. This elusive nature poses a major challenge for exploring its neural basis. While the activity‐silent working memory (ASWM) framework proposes that such latent representations are maintained through transient reconfiguration in intrinsic functional connectivity, tracking these rapid dynamics via functional MRI remains methodologically difficult. Alternatively, investigating the brain's intrinsic functional architecture provides insights into the baseline “neural scaffolding” that predisposes individuals to successful passive WM. To explore this, we combined pre‐task resting‐state fMRI, structural MRI, and behavioral data from a sequential change‐detection paradigm in 151 healthy adults. Functional connectivity–behavior associations revealed that individual differences in passive WM performance were significantly reflected by intrinsic connections among large‐scale networks encompassing dorsal attention, control, and sensorimotor. Granger causality analyses further revealed a group‐level temporal dependency pattern linking these functional systems. Furthermore, exploratory structural analyses suggested spatial convergence between uncorrected cortical‐thickness associations and certain functionally identified sensorimotor nodes. Overall, this study adopts an exploratory approach to demonstrate that baseline intrinsic functional architecture—complemented by preliminary structural findings—is significantly associated with individual differences in passive WM.

Keywords: activity‐silent working memory, control network, directed connectivity, functional connectivity

Key Points

  • Passive WM performance is associated with intrinsic functional connectivity among the dorsal attention, control, and sensorimotor subnetworks.

  • Granger causality analyses revealed a group‐level directional temporal dependency pattern from the dorsal attention subnetwork to the control subnetwork and from the control subnetwork to the sensorimotor subnetwork.

  • Exploratory cortical‐thickness associations showed spatial convergence with the functional findings within the sensorimotor subnetwork, although they did not survive whole‐brain correction.

  • These findings provide an exploratory baseline‐scaffolding account of individual differences in passive WM.

Passive working memory performance is associated with intrinsic connectivity among dorsal attention, control, and sensorimotor subnetworks, suggesting a baseline functional scaffolding for individual differences.

graphic file with name HBM-47-e70597-g004.jpg

1. Introduction

Working memory (WM) is a core cognitive system that supports the temporary maintenance and manipulation of information necessary for complex, goal‐directed behavior. Traditionally, WM has been thought to rely on sustained neural activity during delay periods (D'Esposito and Postle 2015). However, this view has been challenged by recent findings suggesting that WM representations can also exist in a “passive” or “activity‐silent” state, which remains behaviorally accessible via reactivation when needed (Beukers et al. 2021).

Emerging studies have suggested that WM is organized hierarchically, encompassing two distinct functional states: an active and a passive state (Muhle‐Karbe et al. 2021). These two states appear to depend on separate storage resources (Li et al. 2021). Prioritized items are maintained in the active state, directly contributing to ongoing decision‐making and behavioral control; whereas non‐prioritized items reside in the passive state, supporting future retrieval demands (Muhle‐Karbe et al. 2021). The active state, which has been the primary focus of traditional WM paradigms, depends on persistent neural spiking and can thus be detected through sustained patterns of brain activation. In contrast, the passive state lacks persistent neural spiking (Sprague et al. 2016), rendering its neural characterization considerably more challenging.

The “activity‐silent” working memory (ASWM) framework provides a dynamic coding account of the neural mechanisms of WM maintenance. It proposes that passive WM traces are maintained through short‐term synaptic plasticity (STSP), which encodes memory traces by modifying intrinsic functional connectivity (Stokes 2015). Although this framework offers a mechanistic explanation for passive WM, directly tracking such rapid, stimulus‐specific neural reconfigurations with functional magnetic resonance imaging (fMRI) remains challenging because of the sluggish nature of the blood‐oxygen‐level‐dependent (BOLD) signal. A complementary approach is therefore to investigate the brain's intrinsic functional architecture, which reflects relatively stable patterns of functional organization that may provide a neural scaffold for passive WM. Building on this view, the present study does not aim to directly examine the transient mechanisms proposed by the ASWM framework. Rather, we exploratorily investigate whether individual differences in passive WM performance are associated with intrinsic functional network organization, with particular emphasis on cognitive control systems that may facilitate the maintenance of latent memory representations while minimizing interference from irrelevant information (Kreither et al. 2022; Liu et al. 2025).

Within this theoretical framework, we hypothesized that individual differences in passive WM performance can be reflected in the intrinsic functional connectivity, with the cognitive control network serving a central role. We employed a sequential change detection task in 151 participants and acquired resting‐state fMRI data, aiming to examine the association between intrinsic functional network organization and individual differences in passive WM performance. In parallel, structural MRI (sMRI) data were analyzed to determine whether these functional associations are constrained by underlying structural architecture. Overall, this study adopts an exploratory approach to investigate potential relationships between intrinsic functional networks and individual differences in passive WM performance, complemented by exploratory structural analyses.

2. Methods

2.1. Participants

We recruited 162 healthy right‐handed undergraduate volunteers from Sichuan Normal University (Chengdu, China). All participants were native Chinese speakers with normal or corrected‐to‐normal vision and no history of neurological, psychiatric, or major medical disorders, substance abuse, or MRI contraindications (e.g., non‐removable metallic implants, claustrophobia, or pregnancy/lactation). To reduce potential physiological and vasoactive influences on the BOLD signal, heavy smokers (> 5 cigarettes/day), excessive caffeine consumers (> 3 cups/day), and individuals taking medications affecting cognitive or vascular function were excluded. Participants were further instructed to refrain from consuming caffeine, alcohol, and nicotine on the day of the experiment, including a minimum abstinence period of 12 h before scanning.

Two individuals were excluded due to missing imaging data, and nine were excluded for excessive head motion during scanning (see Section 2.4, Resting‐state image analysis), yielding a final sample of 151 participants (104 females). All procedures were approved by the Ethics Committee of Sichuan Normal University. Written informed consent was obtained from each participant, and monetary compensation was provided.

2.2. Acquisition of MRI Data

MRI data were acquired at the Brain Imaging Center of Sichuan Normal University using a Siemens Prisma 3.0‐T scanner with a 64‐channel head–neck coil.

Structural imaging (T1‐weighted MPRAGE): 3D acquisition, TR = 2530 ms, TE = 2.98 ms, flip angle = 7°, FOV = 256 × 224 mm2, matrix = 512 × 448, slice thickness = 1 mm, no gap, voxel size = 0.5 × 0.5 × 1 mm3, 192 slices, scan duration = 6 min.

Resting‐state functional imaging (BOLD EPI): TR = 2000 ms, TE = 30 ms, flip angle = 90°, FOV = 224 × 224 mm2, matrix = 112 × 112, slice thickness = 2 mm, slice gap = 0.3 mm, voxel size = 2 × 2 × 2.3 mm3, 62 slices, scan duration = 8 min. During the scan, participants remained awake, kept their heads still, and fixated on a central cross with their eyes open.

2.3. Task Paradigm

After completing the MRI scan, participants exited the scanner and subsequently completed the behavioral testing session. Passive state WM performance was assessed using a sequential change‐detection paradigm adapted from (Zhang et al. 2022). Two memory arrays were presented in sequence and subsequently tested in reverse order. The second array, probed immediately, reflected active state WM. In contrast, the first array required maintenance during the processing of the second array, without sustained neural activity until retrieval, thereby indexing passive state WM.

Behavioral tasks were programmed in E‐Prime 2.0 and displayed on a 19‐in. monitor (60 Hz, 1920 × 1080 resolution). Participants were seated 70 cm from the screen, instructed to fixate centrally and to prioritize accuracy over speed. Before the main experiment, they received detailed instructions and completed practice trials to ensure task comprehension.

Each participant completed three blocks of 36 trials. To discourage verbal recoding, a unique four‐digit number was shown at the beginning of each block, which participants silently rehearsed throughout (Shaffer and Shiffrin 1972). As illustrated in Figure 1, each trial began with a 500‐ms fixation cross, followed by Memory Array 1 (four colored squares, 500 ms). After a 1000‐ms delay, the low‐load Memory Array 2 (two colored squares, 500 ms) was presented, followed by another 1000‐ms delay. Two probes were then presented sequentially: Probe 1 tested Array 2 and Probe 2 tested Array 1. Each probe contained the same number of items as its corresponding array, with one item colored (Figure 1). The probe color either matched or mismatched the corresponding item (50% probability, randomized). Participants indicated by key press whether the probe color matched, with no time pressure (Figure 1).

FIGURE 1.

FIGURE 1

Open in a new tab

The sequential change‐detection paradigm for assessing passive WM performance. The paradigm is adapted from (Zhang et al. 2022). Memory Array 1 and its corresponding Probe 1 tested passive WM, whereas Memory Array 2 was tested immediately and thus engaged active WM. The orientation of Memory Array 2 was either vertical or horizontal, with each configuration appearing an equal number of times.

Task performance in the passive state WM condition was quantified using accuracy (ACC). Specifically, ACC was calculated as the proportion of correct responses to probes testing Memory Array 1. We additionally computed ACC for the active state WM condition (Memory Array 2), in which items were probed immediately after presentation. This active state WM ACC served as a reference measure in subsequent analyses to evaluate the specificity of the associations observed for passive state WM.

2.4. Resting‐State Images Analysis

Prior to functional preprocessing, quality assessment was performed on all participants' functional images using MRIQC v24.1.0 (Esteban et al. 2017; https://mriqc.readthedocs.io; RRID:SCR_022942), and datasets with excessive head motion were excluded. Exclusion criteria were defined as absolute head motion > 2 mm, average frame displacement (FD) > 0.2 mm, or more than 30% of volumes with FD > 0.2 mm (Power et al. 2012). Of the 160 participants, 9 were excluded for meeting one or more of these criteria, resulting in a final sample of 151 participants for subsequent analyses.

MRI image data were preprocessed using fMRIPrep v23.0.2 (Esteban et al. 2019; https://fmriprep.org; RRID:SCR_016216) which incorporates FreeSurfer v7.3.2 (Dale et al. 1999; https://surfer.nmr.mgh.harvard.edu; RRID:SCR_001847). Functional images underwent slice‐timing correction, motion correction, co‐registration to the T1w reference, and normalization to Montreal Neurosciences Institute (MNI) standard space. Automatic removal of motion artifacts using independent component analysis (ICA‐AROMA) was performed on the preprocessed BOLD time‐series in MNI space, following removal of non‐steady state volumes and spatial smoothing with an isotropic, Gaussian kernel of 6 mm full‐width half‐maximum (Pruim et al. 2015). The standard ‘non‐aggressive’ denoising procedure was utilized, which only removes the variance uniquely associated with the noise components, thereby preserving the shared variance with the neural signal and preventing overly aggressive signal loss. Across all participants, the ICA decomposed the data into an average of 37.74 (± 5.60) independent components. Among these, an average of 18.52 (± 5.43) components, corresponding to 48.73% (± 10.21%) of all components, were classified as motion‐related artifacts and subsequently regressed out. The resulting denoised data were used in subsequent functional connectivity analyses.

Functional connectivity was computed using the CONN toolbox version 22.2407 (Whitfield‐Gabrieli and Nieto‐Castanon 2012; https://web.conn‐toolbox.org; RRID:SCR_009550). ROI‐to‐ROI (region of interest) connectivity matrices were derived across the 200 regions of the Schaefer 17‐network parcellation (Schaefer et al. 2018). Fisher‐transformed bivariate correlations from weighted general linear models (GLM) represented connectivity strength between each pair of ROIs. Group‐level analyses were performed using GLM. For each connection, a separate model was estimated with first‐level connectivity values as dependent variables and passive WM ACC as the main predictor, controlling for age, gender and mean framewise displacement (FD). Connection‐level effects were assessed using multivariate parametric statistics with random effects across subjects and covariance estimation across measurements. Cluster‐level inference was performed using Spatial Pairwise Clustering (SPC) (Zalesky et al. 2012), with 1000 residual randomizations, grouping ROIs according to the intrinsic modular structure of the Schaefer parcellations. Results were thresholded using a cluster‐forming connection‐level threshold of p < 0.001 and a cluster‐mass familywise correction of p‐FWE < 0.05.

2.5. Granger Causality Analysis

As an exploratory follow‐up to the initial functional connectivity analysis, we applied Granger Causality (GC) analysis to explore the directionality of the connections among the specific clusters identified in the preceding step. Specifically, GC was employed to assess the asymmetric temporal predictive relationships between these data‐derived regions. It is important to note that GC reflects statistical prediction over time and does not imply true neural causality. Therefore, interpretations based on GC should be considered indicative of directional associations rather than definitive causal relationships.

We conducted GC analyses using the Multivariate Granger Causality (MVGC) Toolbox version 1.3 (Barnett and Seth 2014; https://users.sussex.ac.uk/~lionelb/MVGC; RRID:SCR_015755). A bivariate autoregressive model was estimated for each pair of network clusters, with model parameters fitted using ordinary least squares (OLS) regression. The optimal model order was automatically determined using the Akaike Information Criterion (AIC; maximum order = 10). Statistical significance of GC estimates was assessed at α = 0.05. For each pair of network connections, both ends consisted of a cluster of nodes belonging to the same network. The BOLD signals of all nodes within each cluster were averaged to obtain a representative time series, which was then used in MVGC analyses to infer the directional influence between the two clusters.

To reduce the risk of circularity associated with behavior‐selected node definitions, we additionally performed an atlas‐defined network‐level GC sensitivity analysis. Specifically, representative BOLD time series were extracted from the complete Schaefer atlas‐defined subnetworks corresponding to the functional systems implicated by the FC–behavior association analysis. All parcels within each predefined subnetwork were included. These network‐level time series were then entered into the MVGC model to estimate directional temporal predictive relationships.

To further examine whether individual differences in the strength of the observed directed temporal pattern were associated with passive WM performance, we conducted exploratory GC–behavior association analyses. These analyses focused on pathways that exhibited significant directional asymmetry in the GC analysis, because the aim was to evaluate whether the magnitude of the observed directional temporal dependency was related to behavioral performance. For each examined pathway, a directional asymmetry index was calculated by subtracting the GC estimate in the reverse direction from the GC estimate in the forward direction. Partial correlation analyses were then conducted between these directional asymmetry indices and passive WM accuracy, controlling for age, gender, and mean FD.

2.6. Cortical Thickness‐Behavior Association

Given that functional neural activity is constrained by cortical structure, we further explored whether regional cortical thickness was associated with individual differences in passive WM performance, following the analysis of functional connectivity. Cortical thickness was used as a structural measure to perform exploratory correlations with passive working memory ACC. Cortical reconstruction and thickness estimation were performed using the FreeSurfer v7.3.2 pipeline embedded within fMRIPrep v23.0.2. This process included intensity non‐uniformity correction, skull stripping, tissue segmentation, surface reconstruction, and thickness estimation. The resulting cortical thickness maps were generated in the fsLR surface space with a 91 k vertex resolution. For regional analysis, mean cortical thickness values were extracted for each parcel using the Schaefer 17‐network atlas with 200 parcels (Schaefer et al. 2018).

To examine the relationship between cortical thickness and behavioral performance, an exploration whole‐brain partial correlation analysis was conducted. Partial correlations were computed between the cortical thickness of all 200 parcels (based on the Schaefer atlas) and the accuracy of the passive working memory. Age and sex were included as covariates to control for potential confounding effects. Statistical significance was assessed using an FDR correction at α = 0.05. In addition, for exploratory purposes, we also examined findings that did not survive FDR correction but exhibited spatial convergence with the functionally identified regions.

3. Results

3.1. Demographics

A total of 151 participants were included in the subsequent analyses, of whom 104 were female (68.9%). The mean age was 20.1 years (range: 18–27 years, SD = 2.19).

3.2. Functional Connectivity‐Behavior Association Analysis

Behavioral performance, indexed by accuracy for passive WM obtained from the sequential change‐detection paradigm, had a mean of 0.677 (SD = 0.095) and ranged from 0.44 to 0.94. No data points fell beyond ±3 SDs (Figure 2).

FIGURE 2.

FIGURE 2

Open in a new tab

Behavioral distribution, functional connectivity, and directed connectivity. (A) Distribution of behavioral accuracy for passive WM obtained from the sequential change‐detection paradigm. The color boundaries represent the first and third quartiles. (B) Functional connections significantly associated with passive WM, identified using GLM analysis. (C) Results of the Granger causality analysis showing the group‐level directional temporal dependency pattern among the networks. CN, control network; DAN, dorsal attention network; DMN, default mode network; LN, limbic network; SMN, sensorimotor network; SN, salience network; TempPar, temporo‐parietal network; VisCent, central visual network; VisPeri, peripheral visual network. The suffixes a, b, and c denote distinct subnetworks within each large‐scale network.

The GLM analysis revealed three significant clusters of functional connections (Figure 2): within the dorsal attention a subnetwork (DANa, p‐FWE = 0.030), between the DANa and control a subnetwork (CNa, p‐FWE = 0.047), and between the CNa and somatomotor a subnetwork (SMNa, p‐FWE = 0.019). Statistical significance was determined using SPC, with a connection‐level threshold of p_uncorrected < 0.001 and a cluster‐level threshold of p_FWE < 0.05. Please refer to Table 1 for detailed statistical results and Figure S1 for the precise spatial localization of these cortical regions.

TABLE 1.

Significant functional connections associated with passive WM performance.

Cluster Cluster‐wise FWE‐corrected p Connections (ROI‐ROI) MNI coordinates (x, y, z) t
1 0.019 R CNa IPS 2—R SMNa 7 (46, −38, 49) (29, −34, 65) 3.79
R CNa IPS 2—R SMNa 8 (46, −38, 49) (22, −9, 67) 3.77
R CNa IPS 2 – R SMNa 5 (46, −38, 49) (32, −40, 64) 3.62
R CNa IPS 2 – R SMNa 6 (46, −38, 49) (33, −21, 65) 3.61
R CNa IPS 1 – R SMNa 5 (37, −63, 47) (32, −40, 64) 3.57
R CNa IPS 2 – R SMNa 4 (46, −38, 49) (40, −24, 57) 3.5
R CNa IPS 1 – R SMNa 9 (37, −63, 47) (10, −39, 69) 3.45
R CNa IPS 2 – R SMNa 2 (46, −38, 49) (47, −11, 48) 3.41
R CNa PFCd 1 – R SMNa 5 (26, 7, 58) (32, −40, 64) 3.40
R CNa IPS 2 – R SMNa 3 (46, −38, 49) (7, −11, 51) 3.36
2 0.030 L DANa SPL 2 – R DANa SPL 3 (−26, −70, 38) (34, −48, 51) 4.32
L DANa SPL 3 – R DANa SPL 3 (−54, −27, 42) (34, −48, 51) 3.80
L DANa SPL 1 – R DANa SPL 3 (−57, −60, −1) (34, −48, 51) 3.74
L DANa SPL 2 – R DANa SPL 2 (−26, −70, 38) (15, −73, 53) 3.71
L DANa SPL 3 – R DANa SPL 4 (−54, −27, 42) (26, −61, 58) 3.49
L DANa SPL 2 – R DANa SPL 4 (−26, −70, 38) (26, −61, 58) 3.37
3 0.047 L CNa PFCl 2 – L DANa SPL 1 (−48, 6, 29) (−57, −60, −1) 4.07
L CNa PFCl 2 – L DANa SPL 2 (−48, 6, 29) (−26, −70, 38) 3.69
L CNa PFCl 2 – L DANa SPL 3 (−48, 6, 29) (−54, −27, 42) 3.46
Open in a new tab

Note: This table lists the significant functional connections related to passive WM as identified by the GLM analysis. The correction method used spatial pairwise clustering (SPC). All reported connections survived at the connection‐level uncorrected threshold of p < 0.001. MNI (Montreal Neurological Institute) coordinates are presented in millimeters (mm). For each connection, the two sets of coordinates correspond to the first and second ROIs listed in the “Connections (ROI–ROI)” column, respectively. The suffixes a, b, and c denote distinct subnetworks within each large‐scale network.

Abbreviations: CN, control network; DAN, dorsal attention network; IPS, intraparietal sulcus; PFCd, dorsal prefrontal cortex; PFCl, lateral prefrontal cortex; R/L, right/left hemisphere; SMN, sensorimotor network; SPL, superior parietal lobule.

For active WM, accuracy averaged 0.837 (SD = 0.070), ranging from 0.65 to 0.98. The GLM analysis identified no significant clusters of functional connections. Statistical significance was also determined using SPC, with a connection‐level threshold of p_uncorrected < 0.001 and a cluster‐level threshold of p_FWE < 0.05.

3.3. Granger Analysis

In the functional connectivity‐behavior association analysis, the three identified connection clusters corresponded to: (1) intra‐network connections among specific nodes within the DANa, (2) inter‐network connections between specific nodes of the DANa and CNa, and (3) inter‐network connections between specific nodes of the CNa and SMNa. In the subsequent GC analysis, we examined the directionality of the connections within these three data‐derived clusters. Within‐network interactions between DANa subclusters showed no significant directionality (t = 1.49, p = 0.137, CI = [−0.002, 0.015]), with the average proportion of explained variance (PEV) being 4.2% and 3.6%, respectively. Between‐network analyses further revealed that DANa subcluster exerted a significant directed temporal influence on CNa subcluster (t = 7.05, p = 5.99 × 10−11, CI = [0.036, 0.065]). The average PEV from DANa subcluster to CNa subcluster was 8.1%, compared with 3.6% from CNa subcluster to DANa subcluster. CNa subcluster showed a significantly stronger directed temporal influence on SMNa subcluster (t = 4.65, p = 7.14 × 10−6, CI = [0.012, 0.030]). The average PEV from CNa subcluster to SMNa subcluster was 5%, while in the reverse direction it was 3.0%. CI represents 95% confidence interval. Degrees of freedom are 150.

By applying the GC model to these predefined, comprehensive network boundaries rather than customized clusters, the network‐level sensitivity analysis revealed a consistent directional pattern. Specifically, DANa exerted a significant directed temporal influence on CNa (t = 2.960, p = 0.004, CI = [0.005, 0.023]). The average PEV from DANa to CNa was 5.3%, compared with 4.0% from CNa to DANa. Similarly, CNa showed a significantly stronger directed temporal influence on SMNa than the reverse direction (t = 2.908, p = 0.004, CI = [0.004, 0.022]). The average PEV from CNa to SMNa was 4.7%, while in the reverse direction it was 3.5%. CI represents 95% confidence interval.

To examine whether individual differences in the strength of directed temporal asymmetry were associated with passive WM performance, we conducted partial correlation analyses between the GC directional asymmetry indices and passive WM accuracy, controlling for age, gender, and mean FD. For the GC analysis based on data‐driven clusters, passive WM accuracy was not significantly associated with the directional asymmetry index for the DANa subcluster → CNa subcluster pathway (r = 0.005, p = 0.950) or the CNa subcluster → SMNa subcluster pathway (r = −0.081, p = 0.325). Interactions between the DANa subclusters were not included in the exploratory GC–behavior association analysis because they did not show significant directional asymmetry in the GC analysis and therefore did not yield an interpretable directional asymmetry index. Similarly, for the atlas‐defined network‐level GC analysis, passive WM accuracy was not significantly associated with the directional asymmetry index for the DANa → CNa pathway (r = −0.073, p = 0.379) or the CNa → SMNa pathway (r = −0.089, p = 0.282).

3.4. Cortical Thickness‐Behavior Association

In the exploratory whole‐brain structural analysis (across all 200 parcels), partial correlation analyses controlling for age and sex (degrees of freedom = 147) revealed that no single region survived the whole‐brain FDR correction (the smallest corrected p‐value was 0.465). However, examining the uncorrected results revealed an exploratory spatial convergence: out of the entire brain, two of the top three regions with the lowest p‐values were the previously identified right SMNa 4 (r = 0.221, p = 0.007, p‐FDR = 0.465) and right SMNa 8 (r = −0.220, p = 0.007, p‐FDR = 0.465), alongside right LimbicA TempPole 2 (r = −0.221, p = 0.007, p‐FDR = 0.465). The anatomical locations of these two SMN subregions, which showed spatial convergence with the functional findings, correspond to the right postcentral gyrus (PoCG) and right precentral gyrus (PreCG), respectively, according to the automated anatomical atlas 3 (Rolls et al. 2020). Beyond the regions reported above, no other functionally identified regions demonstrated a cortical thickness–behavior association with an uncorrected p value below 0.05. It is worth noting that these anatomical connections are presented as exploratory observations, suggesting a spatial convergence with functional findings, and therefore should be interpreted with caution. Please refer to Figure 3 for spatial visualizations and Table 2 for detailed exploratory statistical results.

FIGURE 3.

FIGURE 3

Open in a new tab

Partial correlation results between cortical thickness and passive WM accuracy. Panels A and B correspond to the right SMNa‐4 and SMNa‐8 regions, respectively, within the Schaefer 17‐network, 200‐parcellations. Their anatomical locations correspond to the right postcentral gyrus (PoCG) and right precentral gyrus (PreCG), respectively, as defined by the Automated Anatomical Labeling Atlas 3 (AAL3).

TABLE 2.

Exploratory whole‐brain analysis of correlations between regional cortical thickness and passive working memory accuracy.

Cortical regions MNI coordinates (x, y, z) r Uncorrected p p‐FDR
R SMNa 4 (40, −24, 57) 0.221 0.007 0.465
R SMNa 8 (22, −9, 67) −0.220 0.007 0.465
R LNa TempPole 2 (47, −12, −35) −0.221 0.007 0.465
L DMNa PFCm 3 (−6, 44, 7) −0.211 0.010 0.494
R SMNa 10 (6, −23, 69) −0.179 0.029 0.837
L DMNa PCC 1 (−5, −55, 27) −0.170 0.038 0.837
L CNa PFClv 1 (−42, 40, 16) −0.165 0.045 0.837
Open in a new tab

Note: Results are derived from a whole‐brain exploratory analysis across all 200 parcels of the Schaefer atlas. Regions listed are those exhibiting the top spatial peaks at an uncorrected threshold of p < 0.05. No regions survived the whole‐brain FDR correction. Therefore, these anatomical associations are presented as exploratory evidence demonstrating a spatial convergence with the functional findings and should be interpreted with caution. MNI (Montreal Neurological Institute) coordinates are presented in millimeters (mm). The suffix “a” denotes the specific subnetwork within each large‐scale network.

Abbreviations: CN, control network; DMN, default mode network; LN, limbic network; PCC, posterior cingulate cortex; PFClv, ventral lateral prefrontal cortex; PFCm, medial prefrontal; SMN, sensorimotor network.

4. Discussion

In this exploratory study, we combined resting‐state functional connectivity, directed connectivity analysis, and cortical morphometry to investigate the intrinsic neural architecture underlying individual differences in passive WM. The results revealed that individual differences in passive WM performance were significantly reflected by functional communication among three key cortical networks: the DANa, the CNa, and the SMNa. In contrast, no functional connections showed significant associations with active WM performance. At the population level, the Granger causality analysis revealed a directional temporal dependency pattern among the DANa, CNa, and SMNa. However, because the strength of this pattern was not significantly associated with passive WM performance, it should be interpreted as a descriptive feature of the resting‐state functional architecture rather than as a mechanism directly explaining behavioral performance. At the structural level, exploratory analyses indicated that cortical thickness within the SMNa showed spatial convergence with the functional findings. In summary, this study suggests that cross‐network interactions among attention, control, and sensorimotor systems constitute an intrinsic functional architecture associated with passive WM performance. In addition, exploratory spatial convergence within the sensorimotor cortex provides complementary anatomical context for these functional associations.

4.1. Functional Interactions and Temporal Dependencies Across Functional Networks

Our study identified that individual differences in passive WM are intrinsically linked to three large‐scale cortical networks: the DANa, CNa and SMNa. Within the ASWM framework, we propose that the intrinsic organization of these networks may provide a baseline functional context associated with individual differences in passive WM performance.

The functional systems identified in the present study are broadly consistent with known cognitive roles of these networks. Specifically, the DANa is primarily responsible for top‐down attentional monitoring and selection (Majerus et al. 2018), with the Superior Parietal Lobule (SPL) implicated in our study serving as a key region for spatial orienting and maintenance (Vandenberghe et al. 2001). The regions identified also encompassed core nodes of the CNa, including the lateral prefrontal cortex (PFC), dorsal PFC, and intraparietal sulcus (IPS). As the seat of cognitive control, the CNa contributes to goal maintenance, task updating, planning, and inhibition (Schaefer et al. 2018). In the context of passive WM, these processes may be relevant to maintaining task goals and reducing interference, although the present resting‐state design does not directly test these task‐specific operations. The SMNa is involved in sensorimotor representation and action‐related processing (Gordon et al. 2023), which may provide a plausible functional context for its association with passive WM performance.

The GC analysis further revealed a group‐level pattern of directional temporal dependencies among these functional systems. This pattern is compatible with an ordered relationship among attention‐, control‐, and sensorimotor‐related systems. However, these GC findings should not be interpreted as evidence for a causal hierarchy or as a mechanism directly explaining individual differences in passive WM performance. First, GC reflects statistical temporal prediction rather than biological causality. Second, our exploratory GC–behavior analyses showed that individual differences in GC directional asymmetry were not significantly associated with passive WM accuracy. Therefore, the GC results are best understood as a descriptive characterization of the resting‐state temporal organization among the implicated networks, rather than evidence that stronger directed temporal dependencies lead to better behavioral performance.

Together, the present functional findings suggest that passive WM performance is associated with intrinsic functional communication among attention, control, and sensorimotor networks. Notably, the involvement of the CNa indicates that the control network, although traditionally emphasized in active WM (Miller et al. 2018), may also contribute to the intrinsic functional architecture associated with passive WM performance. The GC results provide complementary descriptive information regarding the group‐level temporal organization of these systems, but their behavioral relevance and mechanistic significance remain to be established in future studies.

4.2. Structural–Behavioral Associations

The exploratory whole‐brain cortical‐thickness analysis did not identify any region that survived FDR correction. Therefore, the structural findings should be interpreted with caution. At the uncorrected level, however, two sensorimotor parcels that overlapped spatially with the functional findings were among the regions with the lowest p‐values. We report this pattern as an exploratory spatial convergence between the functional and structural observations, rather than as evidence for an integrated structural–functional mechanism.

The structural associations within the sensorimotor network were not consistent. Specifically, cortical thickness in the PreCG and PoCG showed opposite uncorrected associations with passive WM performance. This pattern complicates interpretation and argues against a simple structural account. At the same time, both regions were involved in functional connections with the CNa IPS that showed positive associations with passive WM performance. Thus, whereas the functional findings showed a relatively consistent association with behavior, the cortical‐thickness findings appeared more regionally heterogeneous.

One possible explanation is that cortical thickness reflects multiple neurodevelopmental and microstructural factors whose behavioral implications may vary across cortical regions (Choi et al. 2024; Ehrlich et al. 2012). The PreCG and PoCG also contribute to partially distinct sensorimotor processes, with the former more closely related to motor preparation and execution (Porro et al. 1996) and the latter more closely related to somatosensory processing (Alahmadi 2024). Accordingly, the opposite directions of the cortical‐thickness associations may reflect regional heterogeneity within the sensorimotor cortex rather than genuinely opposing neural mechanisms. However, because these structural associations were observed only at an uncorrected threshold, this interpretation remains speculative. Future studies with larger samples, multimodal structural measures, and formal cross‐modal models are needed to determine whether and how sensorimotor cortical structure contributes to individual differences in passive WM performance.

4.3. Limitations

Several limitations should be acknowledged. First, and most fundamentally, the study used resting‐state fMRI data acquired prior to the behavioral task, and the observed connectivity patterns reflect baseline neural scaffolding associated with passive WM, rather than direct evidence of passive WM characteristics. Second, although Granger causality offers insight into directed dependencies between brain regions, it does not establish true neural causality and may be affected by factors such as hemodynamic delays, sampling rates, or unmeasured confounds. Third, the current paradigm focused exclusively on visual–spatial WM; thereby, it remains to be determined whether these findings generalize to other WM domains, such as verbal or auditory memory. Fourth, our structural findings did not survive whole‐brain FDR correction, and the study lacks a deeper multimodal integration. Consequently, these exploratory anatomical associations must be interpreted with caution, and future research with larger cohorts is needed to establish a more definitive cross‐modal synthesis.

5. Conclusion

In conclusion, this exploratory study demonstrates that individual differences in passive working memory can be reflected by the brain's intrinsic functional architecture. Specifically, this neural scaffolding comprises functional interactions among the dorsal attention, cognitive control, and sensorimotor networks. At the population level, GC analyses revealed a directional temporal dependency pattern among these systems, linking attention‐, control‐, and sensorimotor‐related networks. Furthermore, preliminary structural associations showed exploratory spatial convergence within the sensorimotor cortex, providing tentative anatomical context for the functional findings, although these results did not survive whole‐brain correction. Together, rather than directly capturing transient task‐induced states, these findings suggest that passive WM is associated with intrinsic functional and structural features of large‐scale cortical systems, providing a preliminary baseline‐scaffolding account within the ASWM framework.

Funding

This work was supported by the General Project of Humanities and Social Sciences Research of Chinese Ministry of Education (25YJA190009).

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Figure S1: Functional connections significantly associated with passive WM.

HBM-47-e70597-s001.docx (267.4KB, docx)

Acknowledgments

The authors have nothing to report.

Data Availability Statement

All analyses were conducted using publicly available neuroimaging toolkits: MRIQC (v24.1.0), fMRIPrep (v23.0.2), the CONN Toolbox (v22.2407), and the Multivariate Granger Causality (MVGC) Toolbox (v1.3). MRIQC and fMRIPrep are available from the official Docker repositories (https://hub.docker.com/r/nipreps/mriqc; https://hub.docker.com/r/nipreps/fmriprep), the CONN Toolbox is available from the NITRC platform (https://www.nitrc.org/projects/conn), and the MVGC Toolbox from GitHub (https://github.com/lcbarnett/MVGC1). The data supporting the findings of this study are available from the corresponding author upon reasonable request.

References

  1. Alahmadi, A. A. S. 2024. “Beyond Boundaries: Investigating Shared and Divergent Connectivity in the Pre−/Postcentral Gyri and Supplementary Motor Area.” Neuroreport 35, no. 5: 283–290. 10.1097/WNR.0000000000002011. [DOI] [PubMed] [Google Scholar]
  2. Barnett, L. , and Seth A. K.. 2014. “The MVGC Multivariate Granger Causality Toolbox: A New Approach to Granger‐Causal Inference.” Journal of Neuroscience Methods 223: 50–68. 10.1016/j.jneumeth.2013.10.018. [DOI] [PubMed] [Google Scholar]
  3. Beukers, A. O. , Buschman T. J., Cohen J. D., and Norman K. A.. 2021. “Is Activity Silent Working Memory Simply Episodic Memory?” Trends in Cognitive Sciences 25, no. 4: 284–293. 10.1016/j.tics.2021.01.003. [DOI] [PubMed] [Google Scholar]
  4. Choi, U. S. , Shim S. Y., Cho H. J., and Jeong H.. 2024. “Association Between Cortical Thickness and Cognitive Ability in Very Preterm School‐Age Children.” Scientific Reports 14, no. 1: 2424. 10.1038/s41598-024-52576-5. [DOI] [PMC free article] [PubMed] [Google Scholar]
  5. Dale, A. M. , Fischl B., and Sereno M. I.. 1999. “Cortical Surface‐Based Analysis. I. Segmentation and Surface Reconstruction.” NeuroImage 9, no. 2: 179–194. 10.1006/nimg.1998.0395. [DOI] [PubMed] [Google Scholar]
  6. D'Esposito, M. , and Postle B. R.. 2015. “The Cognitive Neuroscience of Working Memory.” Annual Review of Psychology 66: 115–142. 10.1146/annurev-psych-010814-015031. [DOI] [PMC free article] [PubMed] [Google Scholar]
  7. Ehrlich, S. , Brauns S., Yendiki A., et al. 2012. “Associations of Cortical Thickness and Cognition in Patients With Schizophrenia and Healthy Controls.” Schizophrenia Bulletin 38, no. 5: 1050–1062. 10.1093/schbul/sbr018. [DOI] [PMC free article] [PubMed] [Google Scholar]
  8. Esteban, O. , Birman D., Schaer M., Koyejo O. O., Poldrack R. A., and Gorgolewski K. J.. 2017. “MRIQC: Advancing the Automatic Prediction of Image Quality in MRI From Unseen Sites.” PLoS One 12, no. 9: e0184661. 10.1371/journal.pone.0184661. [DOI] [PMC free article] [PubMed] [Google Scholar]
  9. Esteban, O. , Markiewicz C. J., Blair R. W., et al. 2019. “fMRIPrep: A Robust Preprocessing Pipeline for Functional MRI.” Nature Methods 16, no. 1: 111–116. 10.1038/s41592-018-0235-4. [DOI] [PMC free article] [PubMed] [Google Scholar]
  10. Gordon, E. M. , Chauvin R. J., Van A. N., et al. 2023. “A Somato‐Cognitive Action Network Alternates With Effector Regions in Motor Cortex.” Nature 617, no. 7960: 351–359. 10.1038/s41586-023-05964-2. [DOI] [PMC free article] [PubMed] [Google Scholar]
  11. Kreither, J. , Papaioannou O., and Luck S. J.. 2022. “Active Working Memory and Simple Cognitive Operations.” Journal of Cognitive Neuroscience 34, no. 2: 313–331. 10.1162/jocn_a_01791. [DOI] [PMC free article] [PubMed] [Google Scholar]
  12. Li, Z. , Liang T., and Liu Q.. 2021. “The Storage Resources of the Active and Passive States Are Independent in Visual Working Memory.” Cognition 217: 104911. 10.1016/j.cognition.2021.104911. [DOI] [PubMed] [Google Scholar]
  13. Liu, Q. , Wu Y., Yu Q., Yuan Z., Ren G., and Li Z.. 2025. “Offline Memory Representations Have Robustness Superiority in Visual Working Memory.” Acta Psychologica Sinica 57, no. 5: 739. 10.3724/sp.J.1041.2025.0739. [DOI] [Google Scholar]
  14. Majerus, S. , Peters F., Bouffier M., Cowan N., and Phillips C.. 2018. “The Dorsal Attention Network Reflects Both Encoding Load and Top‐Down Control During Working Memory.” Journal of Cognitive Neuroscience 30, no. 2: 144–159. 10.1162/jocn_a_01195. [DOI] [PubMed] [Google Scholar]
  15. Miller, E. K. , Lundqvist M., and Bastos A. M.. 2018. “Working Memory 2.0.” Neuron 100, no. 2: 463–475. 10.1016/j.neuron.2018.09.023. [DOI] [PMC free article] [PubMed] [Google Scholar]
  16. Muhle‐Karbe, P. S. , Myers N. E., and Stokes M. G.. 2021. “A Hierarchy of Functional States in Working Memory.” Journal of Neuroscience 41, no. 20: 4461–4475. 10.1523/JNEUROSCI.3104-20.2021. [DOI] [PMC free article] [PubMed] [Google Scholar]
  17. Porro, C. A. , Francescato M. P., Cettolo V., et al. 1996. “Primary Motor and Sensory Cortex Activation During Motor Performance and Motor Imagery: A Functional Magnetic Resonance Imaging Study.” Journal of Neuroscience 16, no. 23: 7688–7698. 10.1523/JNEUROSCI.16-23-07688.1996. [DOI] [PMC free article] [PubMed] [Google Scholar]
  18. Power, J. D. , Barnes K. A., Snyder A. Z., Schlaggar B. L., and Petersen S. E.. 2012. “Spurious but Systematic Correlations in Functional Connectivity MRI Networks Arise From Subject Motion.” NeuroImage 59, no. 3: 2142–2154. 10.1016/j.neuroimage.2011.10.018. [DOI] [PMC free article] [PubMed] [Google Scholar]
  19. Pruim, R. H. R. , Mennes M., van Rooij D., Llera A., Buitelaar J. K., and Beckmann C. F.. 2015. “ICA‐AROMA: A Robust ICA‐Based Strategy for Removing Motion Artifacts From fMRI Data.” NeuroImage 112: 267–277. 10.1016/j.neuroimage.2015.02.064. [DOI] [PubMed] [Google Scholar]
  20. Rolls, E. T. , Huang C. C., Lin C. P., Feng J., and Joliot M.. 2020. “Automated Anatomical Labelling Atlas 3.” NeuroImage 206: 116189. 10.1016/j.neuroimage.2019.116189. [DOI] [PubMed] [Google Scholar]
  21. Schaefer, A. , Kong R., Gordon E. M., et al. 2018. “Local‐Global Parcellation of the Human Cerebral Cortex From Intrinsic Functional Connectivity MRI.” Cerebral Cortex 28, no. 9: 3095–3114. 10.1093/cercor/bhx179. [DOI] [PMC free article] [PubMed] [Google Scholar]
  22. Shaffer, W. O. , and Shiffrin R. M.. 1972. “Rehearsal and Storage of Visual Information.” Journal of Experimental Psychology 92, no. 2: 292–296. 10.1037/h0032076. [DOI] [PubMed] [Google Scholar]
  23. Sprague, T. C. , Ester E. F., and Serences J. T.. 2016. “Restoring Latent Visual Working Memory Representations in Human Cortex.” Neuron 91, no. 3: 694–707. 10.1016/j.neuron.2016.07.006. [DOI] [PMC free article] [PubMed] [Google Scholar]
  24. Stokes, M. G. 2015. “'Activity‐Silent' Working Memory in Prefrontal Cortex: A Dynamic Coding Framework.” Trends in Cognitive Sciences 19, no. 7: 394–405. 10.1016/j.tics.2015.05.004. [DOI] [PMC free article] [PubMed] [Google Scholar]
  25. Vandenberghe, R. , Gitelman D. R., Parrish T. B., and Mesulam M. M.. 2001. “Functional Specificity of Superior Parietal Mediation of Spatial Shifting.” NeuroImage 14, no. 3: 661–673. 10.1006/nimg.2001.0860. [DOI] [PubMed] [Google Scholar]
  26. Whitfield‐Gabrieli, S. , and Nieto‐Castanon A.. 2012. “Conn: A Functional Connectivity Toolbox for Correlated and Anticorrelated Brain Networks.” Brain Connectivity 2, no. 3: 125–141. 10.1089/brain.2012.0073. [DOI] [PubMed] [Google Scholar]
  27. Zalesky, A. , Fornito A., Egan G. F., Pantelis C., and Bullmore E. T.. 2012. “The Relationship Between Regional and Inter‐Regional Functional Connectivity Deficits in Schizophrenia.” Human Brain Mapping 33, no. 11: 2535–2549. 10.1002/hbm.21379. [DOI] [PMC free article] [PubMed] [Google Scholar]
  28. Zhang, J. , Ye C., Sun H. J., et al. 2022. “The Passive State: A Protective Mechanism for Information in Working Memory Tasks.” Journal of Experimental Psychology. Learning, Memory, and Cognition 48, no. 9: 1235–1248. 10.1037/xlm0001092. [DOI] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Figure S1: Functional connections significantly associated with passive WM.

HBM-47-e70597-s001.docx (267.4KB, docx)

Data Availability Statement

All analyses were conducted using publicly available neuroimaging toolkits: MRIQC (v24.1.0), fMRIPrep (v23.0.2), the CONN Toolbox (v22.2407), and the Multivariate Granger Causality (MVGC) Toolbox (v1.3). MRIQC and fMRIPrep are available from the official Docker repositories (https://hub.docker.com/r/nipreps/mriqc; https://hub.docker.com/r/nipreps/fmriprep), the CONN Toolbox is available from the NITRC platform (https://www.nitrc.org/projects/conn), and the MVGC Toolbox from GitHub (https://github.com/lcbarnett/MVGC1). The data supporting the findings of this study are available from the corresponding author upon reasonable request.

Articles from Human Brain Mapping are provided here courtesy of Wiley

RESOURCES