Exploring Neural Mechanisms Underlying Error-related Impairments in Active Working Memory Suggests an Adaptive Shielding of Contents during Cognitive Control

Abstract

Goal-directed behavior relies on cognitive flexibility—the ability to rapidly adapt ongoing thoughts and behaviors while preserving task-relevant information. The performance monitoring system optimizes such behavior by detecting and evaluating errors, whereas the working memory (WM) system maintains relevant information and protects it from interference. We previously showed that motor errors impaired active WM maintenance (Error-Related Impairment of Active working Memory [ERIAM]; [Wessel, J. R., Jiang, J., & Stolley, J. J. Action errors impair active working memory maintenance. Journal of Experimental Psychology: General, 151, 1325–1340, 2022]). Here, we investigated the source of ERIAM by tracking the contralateral delay activity (CDA), a neural marker of visual WM maintenance, throughout the error-making process. Forty-two participants maintained visual information in WM while performing a flanker task during the delay period. Consistent with prior results, a significant ERIAM effect occurred. We observed the canonical CDA before flanker task onset, and this CDA declined after the flanker task began. Following the flanker response, however, we observed a sustained contralateral positivity (CDAp). Critically, CDA amplitudes did not differ between motor correct and error trials before the flanker task, ruling out a general performance deficit. The CDA was also unaffected immediately after flanker onset, ruling out a perceptual interference explanation. Neural differences emerged only after the flanker response, supporting a genuinely error-related origin of the ERIAM effect. Importantly, CDAp was larger following correct responses than errors, and greater CDAp modulation predicted a smaller ERIAM effect. On the basis of prior work, we interpret CDAp as reflecting disengagement from WM storage, suggesting a possible adaptive shielding strategy. These findings provide clear evidence regarding the timing of the ERIAM effect and suggestive evidence regarding potential mechanisms underlying error-related interference in active WM maintenance. Future work will help determine the functional significance of CDAp and its relationship to WM storage and cognitive control.

INTRODUCTION

Goal-directed behavior relies on cognitive flexibility—the capacity to quickly adapt ongoing thoughts and actions while retaining critical information necessary for successful task performance. The performance monitoring system in the human brain optimizes this flexibility by continuously evaluating actions and detecting errors (Ullsperger, Danielmeier, & Jocham, 2014), whereas the working memory (WM) system maintains information and protects it from interference. A key question is how performance monitoring unfolds and interacts with concurrently maintained information in WM. Traditionally, work on error processing has focused on adaptive, strategic processes that ultimately enhance subsequent behavior (Dutilh et al., 2012; Ullsperger & Szymanowski, 2004; Rabbitt, 1969). However, recent evidence suggests a contrasting perspective: Errors may actually impair some ongoing processes (Wessel, 2018; Ullsperger & Danielmeier, 2016), such as perceptual attention (Decker, Finn, & Duncan, 2020; Beatty, Buzzell, Roberts, & McDonald, 2018; Buzzell, Beatty, Paquette, Roberts, & McDonald, 2017) and motor responses (Guan & Wessel, 2022; Ceccarini et al., 2019; Notebaert et al., 2009). This maladaptive view of error processing is further supported by recent computational modeling works, which suggest that the detrimental effects of errors seem rather inevitable (Adkins, Zhang, & Lee, 2024; Purcell & Kiani, 2016).

To further elucidate the potentially adverse effects of error processing, we recently conducted a study to examine whether motor errors impair active WM maintenance, a phenomenon termed ERIAM (Error-Related Impairment of Active working Memory; Wessel et al., 2022). In a series of experiments, participants retained items in WM while performing a motoric task during the delay period. As expected, action errors committed in the motoric task impaired performance on the final WM probe. This behavioral effect was present across seven different experiments using different motoric tasks and WM contents (Wessel et al., 2022) and has since been replicated by other groups (Brochet, Paubel, & Lemercier, 2025). However, behavioral measurement alone cannot answer when (or how) exactly this impairment emerged during the delay period. Here, we utilized a well-validated ERP component—the contralateral delay activity (CDA)—which purportedly reflects the amount of information actively maintained in WM (Adam, Robison, & Vogel, 2018; Luria, Balaban, Awh, & Vogel, 2016; Vogel & Machizawa, 2004). Thus, reductions in CDA amplitude can be interpreted as information dropping from active WM maintenance (Hakim, Feldmann-Wüstefeld, Awh, & Vogel, 2020). By tracking CDA, we aimed to identify the source and timing of the ERIAM effect.

In the present study, participants performed a continuous delayed estimation task while performing an arrow-flanker task during the delay period. This task was adapted from our previous work on the ERIAM effect (Wessel et al., 2022, Experiment 7), but optimized for EEG. We aimed to test three possible explanations of the ERIAM effect. First, (Hypothesis 1 [H1]) ERIAM may be driven by a trial-wise general performance decrement, causing mistakes on both the motor and WM parts of the same trial (e.g., Cheyne, Solman, Carriere, & Smilek, 2009). This hypothesis predicts that trials with motor errors may also contain subpar encoding of the WM contents, ultimately leading to errors on the probe. Second, (H2) ERIAM could be triggered by perceptual interference from the onset of the flanker task. This hypothesis predicts that on trials with motor errors, the CDA will decrease after the flanker stimuli have onset, but before a motor response has been initiated. Third, (H3) ERIAM may be genuinely driven by error commission in the flanker task. This hypothesis predicts that the CDA will be disrupted following errors in the flanker task compared with CDA following correct responses.

Accordingly, in the present study, we measured CDA across three time windows during the delay period to test these competing explanations. Once we observed a significant difference in CDA between error and correct trials, we further examined whether this change in CDA could predict the ERIAM effect.

METHODS

Participants

Forty-six healthy young adults participated in the experiment. Four participants were excluded from the analysis: two due to noncompletion of the task, one due to chance-level accuracy (50%) in the flanker task, and one due to a technical problem during the experiment, leaving 42 participants in the analysis (mean age = 20.38 years; SD = 2.76; 3 left-handed, 24 female participants). Participants were either paid $15 per hour or received course credit for their participation. All participants had normal or corrected-to-normal visual acuity and normal color vision and did not report taking psychiatric medications. The study was approved by the ethics committee at The University of Iowa (institutional review board #201511709).

Our target sample size was based on both prior empirical findings and task-specific design considerations. For detecting the behavioral ERIAM effect, an a priori power analysis using G*Power 3.1 (Faul, Erdfelder, Buchner, & Lang, 2009) with an effect size of d = 0.34—based on our previous study (Wessel et al., 2022, Experiment 7)—indicated that 55 participants are required to achieve 80% power (one-tailed t test, α = .05). For our ERP component of interest, the CDA, a simulation-based power analysis by Ngiam, Adam, Quirk, Vogel, and Awh (2021) indicated that 25 participants are sufficient to detect the component with 80% power. Given these considerations and the fact that our task design involved a high number of trials per participant and increased task difficulty to ensure a sufficient number of error trials, we targeted a sample size of approximately ∼40 participants.

Apparatus

The experiment took place in a dimly lit experimental room. Experimental stimuli were presented on a 24-in. LCD monitor (model: BenQ XL2420B) at the native resolution of 1920 × 1080 pixels using Psychtoolbox3 (Brainard, 1997) with MATLAB 2017b (MathWorks Inc.) run on a Linux Ubuntu computer. Participants were seated approximately 60 cm from the monitor while a chin rest was not used.

Stimuli

Hybrid WM Task

The main task was designed as a hybrid of the continuous delayed estimation task (Wilken & Ma, 2004; Zhang & Luck, 2008) and the arrow-flanker task based on the paradigm from Experiment 7 in our previous study (Wessel et al., 2022). All stimuli were presented on a gray (red, green, blue = 128 128 128) background. For the WM task, a black fixation dot (6 × 6 pixels) was presented at the center of the screen. Either a left or right arrow (70 × 45 pixels) was used as a cue for the to-be-remembered side. Memory displays showed three colored items (60 × 60 pixels) on each side. Within each side, each item was vertically separated by 180 pixels. The distance from the (invisible) midline to each item was 200–212 pixels. On every trial, the locations of items were randomly jittered by ±10 pixels. Six colors were randomly chosen from a 360-color pool adopted from the previous study (Wessel et al., 2022), with the constraint that no two colors were closer than 30° from each other. A color wheel (radius: 350 pixels, ring thickness: 36 pixels) was presented as a WM probe. For the flanker task (Eriksen & Eriksen, 1974), five up/down arrows were used as stimuli (50 × 45 pixels for each).

Change Detection Task

We used a separate change detection task to measure participants' WM capacity (Vogel & Machizawa, 2004; Luck & Vogel, 1997). We adapted the change detection task paradigm used in our previous work (Adam et al., 2018; experimental code: https://github.com/kcsa/change-detection-task). The colors of squares were chosen from a pool of nine distinct colors: red (red, green, blue = 255 0 0), green (0 255 0), blue (0 0 255), yellow (255 255 0), magenta (255 0 255), cyan (0 255 255), orange (255 128 0), white (255 255 255), and black (1 1 1). Stimuli were shown on a gray background. For Set Size 4 and 6, colors were chosen without replacement from the pool. For Set Size 8, colors were randomly selected from a doubled list of the colors.

Tasks

Hybrid WM Task

The task paradigm is illustrated in Figure 1. Each trial started with a fixation dot presented at the center of the screen for 800 msec, followed by a directional arrow cue (left or right) presented above the fixation dot for 450 msec. This arrow cue indicated the to-be-remembered side. Subsequently, three colored items were presented on both the left and right sides of the hemifield for 300 msec. Participants were instructed to remember the color and the location of items only on the cued side of the display. There was a 600-msec blank delay period with a central fixation dot before the presentation of five vertically aligned arrows for the flanker task (one central “target” arrow, surrounded by four flanker arrows). The five arrow stimuli were presented for 100 msec at the center. On each trial, the flanker arrows pointed in either the same direction (e.g., congruent trials) or the opposite direction (e.g., incongruent trials) as the target arrow (equal probability of congruent vs. incongruent). Participants were asked to respond to the target arrow as quickly as possible using their left hand. They used the middle and index fingers to respond to the two possible arrow directions: the key A for the up arrow and the key Z for the down arrow, respectively. Participants were asked to respond within a response time window, which was adaptively adjusted (min: 330 msec, max: 800 msec) based on their performance to ensure an adequate number of erroneous trials for each participant. If participants failed to respond within a response window, “TOO SLOW!” feedback was displayed at the center for 300 msec. The time between the flanker stimuli offset and the WM probe onset was fixed to 1500 msec. After the delay, the WM probe display was presented, featuring a color wheel with one empty black box located at one of the three color squares presented on the cued side. The color wheel was randomly rotated on each trial to discourage participants from coding colors using locations on the color wheel. Participants were instructed to report the color of probed memory item by clicking the corresponding color on the color wheel within 4000 msec. At the beginning of the WM probe, a mouse cursor was shown at the center of the screen. Once participants moved the cursor from the center of the screen, the current color at the cursor's position was displayed as a colored square at fixation. Following the response or after the maximum time window, a blank screen with a fixation dot was presented for 800 msec and the next trial began. The experimental session consisted of 24 blocks of 32 trials, totaling 768 trials. Between blocks, participants received a summary of their performance in the previous block, including the mean accuracy (%) for the flanker task and the mean angular error for the WM probe. In addition, their cumulative performance was displayed. Cumulative mean accuracy (%) for the WM task was calculated as the percentage of WM trials where participants had an offset error of 60° or less. This WM accuracy metric was solely used to encourage participants to have above 65% accuracy in both tasks. During the task, participants were instructed not to move their eyes from the fixation and to pay attention to the cued side of the items covertly. They were also asked to minimize any misses in both tasks.

Figure 1. .

Task design. The duration of each event is shown in milliseconds. In each trial, participants performed a continuous delayed estimation task combined with an arrow flanker task during the delay period. After the flanker task, a WM probe was presented. The response deadline for the flanker task was adjusted based on the participant's performance whereas the interval between the flanker stimuli offset and the WM probe was 1500 msec. The interval from memory array onset to the WM probe was fixed as 2500 msec (detailed timings time-locked to memory array onset were marked in bold at the bottom of the figure).

Change Detection Task

On each trial, either four, six, or eight colored items were presented on the screen for 250 msec. Participants were asked to remember the color and the locations of the items. After a blank delay period of 1000 msec, a WM probe was presented at one of the remembered locations. Participants were asked to report whether the color was changed or not by pressing the “Z” key for no change and the “/?” key for change. The probability of a change trial was 0.5. Immediately after a response, the next trial began.

Procedure

Participants first completed 192 trials of the change detection task (96 trials of two blocks, 64 trials for each set size of 4, 6, and 8), followed by the main hybrid WM task. Before starting the main task, participants underwent one or two blocks of practice, consisting of 12 trials for one block. Trials in the practice block were the same as trials in the main task, except two things: The target arrow at the center was only shown in the flanker task. In addition, participants were given feedback on responses: For a response made within a response window on the flanker task, they were given “correct” or “incorrect” feedback on the screen. For a response on the WM probe, they were provided with feedback about how far the reported color was from the original color as an angular error (e.g., “35 degrees away”). After participants fully understood the task, they moved on to the main task and completed 768 trials. The total session length was up to 3 hr.

EEG Data Acquisition

EEG data were recorded using a 64-channel active electrode cap connected to an actiCHamp amplifier (Brain Products) following the standard International 10–10 system. Incoming data were filtered (low cutoff = direct current, high cutoff = 140 Hz) and collected at a sampling rate of 500 Hz, with channel impedance kept below 10 kΩ throughout the recording session. Ground and reference channels were placed at AFz and Pz, respectively. HEOG was not separately recorded for tracking saccadic movements; instead, saccadic artifacts were corrected using independent component analysis (ICA) during offline EEG preprocessing. Our decision was referred to previous studies reporting that ICA-based correction preserves the CDA (Merkel et al., 2021; Drisdelle, Aubin, & Jolicoeur, 2017).

EEG Preprocessing

After data acquisition, EEG data were processed using custom MATLAB scripts and the EEGLAB toolbox (Delorme & Makeig, 2004). Individual data were imported into MATLAB and downsampled to 250 Hz. Data were filtered using a least-squares finite impulse response filter (high/low cutoff: 0.1–30 Hz), and the continuous EEG data were epoched from −800 msec to 3000 msec relative to the cue onset (0 msec). To clean the data, we followed a two-step approach. First, nonstereotypical artifacts and noisy channels were visually inspected and removed. Rejected channels were interpolated using spherical spline interpolation, and data were rereferenced offline using an average reference. To enhance ICA decomposition, data were high-pass filtered at 1 Hz before running ICA (Klug & Gramann, 2021; Buzzell et al., 2017). The filtered data were then subjected to temporal Infomax ICA (Bell & Sejnowski, 1995) with an extension to sub-Gaussian sources (Lee, Girolami, & Sejnowski, 1999). The ICA weights were then applied to the original EEG data. Thus, all further analyses were performed on the originally filtered dataset (e.g., 0.1 Hz high-pass filtered). Next, independent components (ICs) representing stereotypical artifact activities (i.e., horizontal/vertical eye movements, electrode, muscular artifacts) were visually inspected together with the automated IC classifier detection algorithm (Pion-Tonachini, Kreutz-Delgado, & Makeig, 2019) and accordingly removed from the data. The cleaned EEG data were then used for analyses.

Behavioral Analyses

WM Capacity Assessment via Change Detection Task

Individual WM capacity was estimated using a separate change detection task. Prior work has demonstrated that WM capacity correlates with CDA amplitude, such that individuals with higher capacity show greater CDA amplitudes (Luria et al., 2016; Luck & Vogel, 2013; Vogel & Machizawa, 2004; Cowan, 2001). It is possible, therefore, that CDA differences between errors and correct responses in the flanker task, as well as the ERIAM effect in general, are affected by one's WM capacity. Thus, accounting for WM capacity was necessary for testing the relationship between the ERIAM effect and CDA changes. In doing so, each participant's WM capacity was computed as K = set size × (hit − false alarm) (Cowan, 2001). K estimates for each set size (4, 6, and 8 items) were then averaged to yield one estimate per participant.

Hybrid WM Task

Only trials in which participants responded to both the flanker and WM tasks were included in analyses. Trials with anticipatory responses in the flanker task (RT < 150 msec) were excluded from the analysis. One participant did not understand the task until the third block, and thus, these trials were not counted as valid trials in the analysis. On the basis of these criteria, the total exclusion rate was 10.36% of trials, on average, per participant (SE = 0.66%).

Flanker task.

We first checked the overall error rate in the flanker task. We then examined the typical interference effects in accuracy (%) and mean RTs (RT in msec), such that incongruent trials would produce poorer accuracy and slower RTs compared with congruent trials. Trials were separated as “motor error” and “motor correct” depending on their responses. For comparisons between error and correct conditions, only incongruent trials were included because errors in congruent trials were rare. The overall number of incongruent error trials for each participant included in the behavioral analysis was 91.62 trials (SE = 5.72, min: 18, max: 161). We also examined posterror processing as conduced in the previous study (Wessel et al., 2022): We examined differences in accuracy and mean RT between postcorrect and posterror trials. To do so, RTs and accuracy of the next correct trials following errors and following correct were averaged respectively for each participant. We also examined the interference effect in posterror and postcorrect trials to see whether errors changed performance in the next trial compared with correct responses.

WM task.

WM performance was calculated as the average offset error for trials in the motor error and motor correct conditions. “Offset error” refers to the average angular difference (in degrees) between the original color and the reported color on a color wheel. The ERIAM was quantified by subtracting the mean offset error in motor correct trials from that in motor error trials, normalized by the mean offset error in motor correct:

$$\text{ERIAM}\text{effect}=\left(\text{motor},,\text{error},,\text{minus},,\text{motor},,\text{correct}\right)/\text{motor}\text{correct}$$

This ERIAM effect was used to examine its relationship with changes in CDA amplitudes (motor error minus motor correct), with only trials included in EEG analyses used for this calculation.

EEG Analyses

Some trials were excluded from EEG analyses due to trial rejections during preprocessing. On average, each participant had 82.74 incongruent error trials available for analysis (SE = 5.27, min:15, max: 154). The ERIAM effect was calculated using only these trials.

EEG Data Segmentation

EEG data were segmented into two time windows: (a) epochs time-locked to WM array onset (T1 epoch) and (b) epochs time-locked to the flanker response onset (T2 epoch). T1 epochs were segmented from −1200 to 2500 msec relative to WM array onset, and the T2 epochs were segmented from −1700 to 800 msec relative to the flanker response onset. ERPs for testing H1–H3 were baselined relative to their respective events. For H1, T1 epochs were baselined using a −200- to 0-msec window relative to WM array onset. For H2, T2 epochs were used, but the baseline was still defined from −200 to 0 msec relative to WM array onset, because CDA was measured before the flanker response. For H3, T2 epochs were baselined using a window from −150 to −50 msec relative to the flanker response onset to exclude the immediate periresponse interval, where response-related activity may begin to diverge between conditions.

CDA Analysis

To compute the CDA, we used the PO7/PO8 electrode pair, where the maximal CDA is expected based on prior work (Roy & Faubert, 2023). CDA was calculated by subtracting activity at the ipsilateral electrode from activity at the contralateral electrode, relative to the cued (to-be-remembered) side. CDA was then averaged across trials to yield one value for motor error and motor correct conditions. To test our three hypotheses, we compared CDA amplitudes between conditions across the following time windows:

For H1, CDA was computed from 500 to 900 msec following WM array onset. This window began 200 msec after array offset and ended at flanker task onset, capturing a period during which CDA typically stabilizes.

For H2, CDA was calculated from flanker task onset to flanker response. Because RTs varied by trial, the window length was trial specific, determined by tracing backward from the flanker response to the flanker onset. Mean CDA was then computed within this window for each trial and averaged across trials separately for motor error and motor correct conditions.

For H3, CDA was examined following the flanker response, up to 800 msec—corresponding to the final portion of the delay period (i.e., the latest possible response window). To determine whether and when postresponse CDA difference emerged, we conducted a cluster-based permutation test across the entire 0- to 800-msec epoch window (two-sided, cluster-level p < .05, 5000 iterations). This data-driven procedure identified temporal clusters showing significant differences between motor error and motor correct trials. Mean CDA was computed within significant cluster intervals for comparisons.

Brain–Behavior Relationships

To assess whether postresponse neural modulation relates to behavioral variability across individuals, we conducted separate brain–behavior analyses.

• (i) CDA-ERIAM relationship

For each participant, the ERIAM effect was calculated as (motor error minus motor correct) / motor correct. To quantify postresponse modulation at the individual level, we computed the mean CDA difference (motor error minus motor correct) across the full 0- to 800-msec postresponse interval. This broader interval was selected to capture overall postresponse modulation while avoiding reliance on a specific fixed or cluster-defined window for correlational analyses. We then tested if individual differences in postresponse CDA modulation predicted the ERIAM effect. Given the strong association between CDA amplitude and WM capacity (Adam et al., 2018; Luria et al., 2016; Vogel & Machizawa, 2004), we computed partial correlations, controlling for individual WM capacity.

• (ii) CDA-WM capacity relationship

We examined the well-established relationship between CDA and WM capacity, which prior work has shown to be positively associated—individuals with greater WM capacity tend to show larger CDA amplitudes (i.e., greater negative deflections; Adam et al., 2018; Luria et al., 2016; Vogel & Machizawa, 2004). For preresponse periods (H1 and H2), CDA was quantified using the hypothesis-defined windows described above. For the postresponse period (H3), mean CDA amplitude across the full 0- to 800-msec interval was used to index overall postresponse modulation. For each window, CDA was averaged across motor error and motor correct trials to obtain a mean CDA value per participant. We then tested correlations between these values and WM capacity separately for each time window.

Statistical Analyses

To compare CDA amplitudes between motor error and motor correct conditions, we used two-tailed paired-samples t tests, with a critical p value of .05. Effect sizes (Cohen's d) were reported for all pairwise comparisons. Bayes factors were also reported for each CDA comparison to quantify the strength of evidence (van Doorn et al., 2021). To address temporal circularity and multiple comparisons in the postresponse period (H3), a cluster-based permutation test was conducted across the full 0- to 800-msec interval (two-sided, cluster level p < .05, 5000 iterations). To assess brain–behavior relationship, we computed Spearman rank-order correlations. For the CDA-ERIAM analysis, we additionally controlled for WM capacity using partial correlation.

RESULTS

Behavioral Results

WM Capacity

Participants' WM capacity was measured through the separate change detection task, across three set sizes 4, 6, and 8. Participants had a mean WM capacity of 2.50 (SE = 0.11) [min: 1.29, max: 4.10].

Flanker Task

Overall, participants had a 15.66% error rate in the flanker task (SE = 0.90%). We found the typical flanker interference effect in the flanker task. First, accuracy was significantly lower in incongruent (M = 72.58%, SE = 1.51) than congruent trials (M = 95.15%, SE = 0.58), t(41) = −15.81, p < .001, d = 2.44. Second, participants made slower correct responses in incongruent (M = 444.59 msec, SE = 6.54) than congruent (M = 384.53 msec, SE = 5.41) trials, t(41) = 16.06, p < .001, d = 2.478. For the following “motor error” and “motor correct” comparisons, only incongruent trials were included in the analyses. First, RTs in error trials were significantly faster (M = 356.82 msec, SE = 6.73) than correct trials (M = 444.59 msec, SE = 6.54), indicating that speeded responses were predominant in the motor error trials (mean difference = 87.77 msec, t(41) = 23.65, p < .001, d = 3.649). We found no difference in accuracy between postcorrect (M = 87.02%, SE = 0.87) and posterror (M = 86.07%, SE = 0.94) trials, t(41) = 1.44, p = .159, d = 0.221. RTs in posterror trials (M = 400.61 msec, SE = 5.92) were significantly faster than postcorrect trials (M = 408.28 msec, SE = 5.36), t(41) = −3.13, p = .003, d = 0.482. Posterror modulation on interference effects were not observed neither in RTs (postcorrect: 57.56 msec [SE = 3.42] vs. posterror: 60.31 msec [SE = 4.18]; t(41) = −0.92, p = .362, d = 0.142) nor in accuracy (postcorrect: −19.88% [SE = 1.45] vs. posterror: −22.42% [SE = 1.45]; t(41) = 1.65, p = .106, d = 0.255).

WM Task

When only correct trials were considered, participants had greater WM errors following responses in incongruent trials (M = 43.74 deg, SE = 2.07) than congruent trials (M = 41.46 deg, SE = 2.00), t(41) = 3.41, p = .001, d = 0.526. Unlike in our earlier study (Wessel et al., 2022), which used a different task design, here, WM performance was affected by the congruency in the flanker task. However, no congruency effect was found for RTs. RTs for WM probe were comparable for congruent trials (M = 1770.37, SE = 37.49) and incongruent trials (M = 1776.96, SE = 38.67), t(41) = −1.06, p = .294, d = 0.164. Next, to examine error-related impairment on WM performance (i.e., ERIAM effect), we compared WM offset errors between motor correct and motor error trials (note, this analysis was only performed for “incongruent” trials). Participants made greater offset errors in the WM probe in the motor error (M = 51.66 deg, SE = 1.78) than motor correct (M = 43.74, SE = 2.07) trials, suggesting that error-related impairment occurred (difference = 7.92; t(41) = −6.29, p < .001, d = 0.97; Figure 2A). The ERIAM effect (M = 0.23, SE = 0.04) indicates that offset errors following motor errors were significantly increased relative to following motor correct responses (one-sample t test, t(41) = 6.06, p < .001, d = 0.935; Figure 2B). Note that the mean difference in WM offset errors between conditions and ERIAM effect included in behavioral and EEG analyses were highly comparable (rs > .959, ps < .001). Detailed summary statistics are provided in Supplemental Table S1. Following errors in the flanker task, participants made faster reports in the WM probes compared with following correct responses (motor error: 1679.86 [SE = 38.00] vs. motor correct: 1776.96 [SE = 38.67], t(41) = −5.49, p < .001, difference = −97.1 msec, d = 0.847).

Figure 2. .

WM performance. (A) WM offset errors between correct (black) and error (red) in the flanker task. (B) The degree of ERIAM effect was calculated as WM offset errors in two conditions.

Because motor errors in the flanker task were associated with both faster flanker RTs and faster WM-probe RTs, one potential concern is that the ERIAM effect may reflect a general speed-accuracy tradeoff or a global reduction in decision threshold rather than an error-specific impairment. To address this possibility, we conducted a complementary analysis binning flanker RTs within participants by median split. A repeated-measures ANOVA on WM offset errors revealed significant main effects of RT Bin and Flanker Task Accuracy. Crucially, however, there was no Flanker Task Accuracy × RT Bin interaction. These findings suggest that although slower responses were associated with larger WM errors overall, the ERIAM effect was present across both fast and slow RT bins. This pattern argues against a simple speed-accuracy tradeoff explanation of the ERIAM effect (see Supplemental Materials; control analysis: speed-accuracy tradeoff).

CDA Results

We analyzed differences in CDA amplitude between motor error and motor correct trials to test three hypotheses for the ERIAM effect. (H1) General performance decrements in error trials, (H2) differential processing of flanker task stimuli, and (H3) error-related processes in the flanker task. To evaluate these hypotheses, we measured CDA during three time windows.

H1: Preflanker Task Window (500–900 msec after the WM Array Onset)

CDA was calculated from 500 to 900 msec following WM array onset (Figure 3A), a time window occurring after WM encoding but before the onset of the flanker task stimuli. CDA was significantly lateralized in both conditions (ps < .001), with no significant difference between motor error trials (M = −0.67, SE = 0.16) and motor correct trials (M = −0.67, SE = 0.15), t(41) = 0.00, p = .998, d = 0; BF₀₁ = 6.00, indicating moderate evidence for no difference. The absence of CDA difference suggests that the ERIAM effect is not attributable to generalized performance deficits or impaired encoding of WM items before the flanker task.

Figure 3. .

CDA dynamics across three time windows testing H1–H3. (A) CDA time-locked to WM array onset. The yellow shading indicates the time window of interest (500–900 msec post array) used to test H1. Arrow denotes flanker task onset. (B) CDA time-locked to the flanker response onset used to test H2. Dashed lines indicate the mean flanker task onset times, traced back from response onset: −356 msec for motor error (red) and −444 msec for motor correct (black). CDA was computed for each trial between task onset and response and then averaged across trials to yield mean CDA for each condition. (C) Lateralized waveforms shown as a sustained contralateral positivity (CDAp) time-locked to flanker response onset for testing H3. The cyan shading marks the significant time window (264–312 msec postresponse; cluster-based permutation test, cluster p < .05) in which CDAp differed between conditions. Shaded error bars represent ± SEM.

H2: Postflanker Task Window 1 (from Flanker Task Onset to Response)

Next, we examined whether CDA differed during the interval between the onset of the flanker task and the response (Figure 3B). On average, this window was longer for motor correct trials (M = 444 msec) than for motor error trials (M = 356 msec), reflecting faster erroneous responses. The CDA in the motor correct trials was significantly different from zero (M = −0.34, SE = 0.15), t(41) = −2.20, p = .034, d = 0.339, whereas CDA in the motor error trials (M = −0.41, SE = 0.22) did not reach significance, t(41) = −1.91, p = .064, d = 0.294. Importantly, there was no significant difference in CDA between the motor error and motor correct conditions, t(41) = 0.43, p = .667, d = 0.067; BF₀₁ = 5.49, indicating moderate evidence for no difference. This suggests that the ERIAM effect is unlikely to stem from early interference in the flanker task.

H3: Postflanker Task Window 2 (after Flanker Response Onset)

Last, we examined CDA amplitudes following the flanker response. Visual inspection revealed that CDA diminished in both motor error and motor correct conditions around the time of the flanker response (Figure 3C), showing that CDA was no longer negatively deflected. By the time of the response, the waveform was no longer negatively deflected and instead showed a sustained contralateral positivity (CDAp). The postresponse CDAp was analyzed using the same contralateral-minus-ipsilateral subtraction as the CDA. A cluster-based permutation test identified a significant window extended from 264 to 312 msec in which CDAp differed between error and correct trials (cluster p < .05, two-sided; highlighted as a cyan window in Figure 3C). CDAp amplitudes were averaged within this window for statistical analysis. In the motor error condition, CDAp was not significantly different from zero (M = 0.31, SE = 0.19), t(41) = 1.61, p = .115, d = 0.249. In contrast, the CDAp in the motor correct condition was significantly greater than zero (M = 0.81, SE = 0.18), t(41) = 4.46, p < .001, d = 0.688. Most notably, CDAp was significantly greater following correct responses than errors, t(41) = 2.34, p = .025, d = 0.360; BF₁₀ = 1.88, indicating only weak evidence for a difference. In summary, the canonical CDA declined in both conditions and transitioned into CDAp. Although CDAp was observed in both conditions, it was significantly larger following correct responses than errors, a pattern contrary to our initial predictions.

Relationship between CDAp and the ERIAM Effect

In the CDA analysis, significant differences between motor error and motor correct conditions emerged only after responses in the flanker task, where the waveform was characterized by a sustained contralateral positivity (CDAp). We next examined whether individual variability in this postresponse CDAp modulation was associated with the behavioral ERIAM effect. To capture broader individual differences in postresponse dynamics, CDAp differences (motor error minus motor correct) were averaged across the entire 0- to 800-msec postresponse interval. Only the trials included in the EEG analyses were used to compute the ERIAM effect, which had a mean value of 0.23 (SE = 0.04). Controlling for individual WM capacity, greater CDAp differences between motor error and motor correct trials predicted smaller ERIAM effects (partial ρ = −0.39, p = .011). It is also worth noting that the ERIAM effect was positively associated with WM capacity when controlling for CDAp differences (partial ρ = 0.446, p = .004), suggesting that individuals with higher WM capacity exhibited larger ERIAM effects.

Relationship between CDA and Individual WM Capacity

Last, we examined the association between CDA amplitude and individual WM capacity across three time windows corresponding to H1–H3. CDA amplitudes were averaged across motor error and motor correct trials within each window. Significant negative correlations between CDA and WM capacity were observed before the flanker response: (i) preflanker window (500–900 msec after WM array onset): (ρ = −0.53, p < . 001; Figure 4A) and (ii) postflanker task onset, preresponse window: (ρ = −0.35, p = .025; Figure 4B). These findings replicate prior work showing that larger CDA amplitudes are associated with higher WM capacity (Adam et al., 2018; Vogel & Machizawa, 2004). In contrast, following flanker response, CDAp amplitude averaged across the entire 0- to 800-msec postresponse interval was not reliably associated with WM capacity (ρ = 0.24, p = .120; Figure 4C). Thus, although the canonical CDA–capacity relationship was preserved before the flanker response, this relationship did not extend to the postresponse period.

Figure 4. .

Relationship between CDA and WM capacity across three time windows corresponding to H1–H3. (A) Preflanker task window: 500–900 msec following memory array onset (used for testing H1). (B) Response preparation window: from flanker task onset to flanker response (used for testing H2). (C) Postresponse window: the entire 0- to 800-msec postresponse window following flanker response (used for testing H3). Each column shows the correlation between CDA amplitude and individual WM capacity for the respective window.

DISCUSSION

In the present study, we investigated the neural source of error-related impairments in active working memory (ERIAM; Wessel et al., 2022). Using the CDA—a neural marker of active WM maintenance during delay periods (Luria et al., 2016; Vogel & Machizawa, 2004)—we aimed to determine the temporal locus of these impairments, thereby identifying the processing stage at which the ERIAM effect emerges. Our findings revealed that the CDA did not differ between motor error and correct trials before the flanker response. Only after responding to the flanker task did we observe differences between conditions, measured as a sustained contralateral positivity (CDAp), suggesting that motor errors directly contribute to the ERIAM effect while ruling out several alternative explanations.

Testing Three Key Time Windows for CDA and ERIAM Effect

Because the ERIAM effect could arise from distinct cognitive mechanisms, we tested CDA dynamics in three key time windows corresponding to three hypotheses:

H1: General Performance Decrement

The ERIAM effect could result from general lapses in WM encoding or maintenance, leading to reduced CDA (Adam et al., 2018; McCollough, Machizawa, & Vogel, 2007) even before the motor error and possibly increasing the likelihood of making motor errors. That is, the WM deficit is not due to the motor error itself, but both are caused by a third factor. However, our findings did not support this explanation: CDA amplitudes before flanker task did not differ between correct and error trials (Figure 3A), indicating comparable WM encoding and maintenance across conditions, even when a motor error followed. Thus, general performance decrements do not explain the ERIAM effect (Hakim, Feldmann-Wüstefeld, Awh, & Vogel, 2021; Hakim et al., 2020; Adam, Mance, Fukuda, & Vogel, 2015; Weissman, Roberts, Visscher, & Woldorff, 2006). Supporting this interpretation, EEG markers of cognitive engagement—frontal theta and lateralized alpha power—also did not differ between conditions (Supplemental Figure S3).

H2: Perceptual Interference from Flanker Task

Another possibility was that flanker task stimuli interfered with WM maintenance, producing the ERIAM effect. Previous studies have shown that stimuli presented during delay periods can impair WM performance (Gresch, Boettcher, Van Ede, & Nobre, 2021; Hakim et al., 2020; Bae & Luck, 2019). Although CDA amplitude declined toward zero after the flanker task onset, CDA amplitudes before responses did not differ between motor error and correct trials (Figure 3B), suggesting that perceptual interference alone cannot explain the ERIAM effect.

H3: Error Commission in Flanker Task

Last, the ERIAM effect may be directly tied to motor error in the flanker task. We found that condition differences emerged only after the motor response, during the period characterized by CDAp, as confirmed by a cluster-based permutation analysis identifying a significant window (264–312 msec; Figure 3C). This indicates error commission—rather than general lapses or perceptual interference—as the driver of WM impairment. In addition, this postresponse CDAp modulation persisted after removal of error-related independent components (Supplemental Figure S4), suggesting that the effect cannot be fully explained by overlapping error-related activity (i.e., the error-related negativity and the error positivity [ERN/Pe]) (Gehring, Goss, Coles, Meyer, & Donchin, 1993; Falkenstein, Hohnsbein, Hoorman, & Blanke, 1991).

Taken together, by identifying the time window in which CDA difference emerged, we provide direct evidence regarding the timing and source of the ERIAM effect. Building on prior behavioral work (Wessel et al., 2022), the current study establishes that error commission itself alters active WM maintenance, offering new insights into how cognitive control and memory systems interact (Watanabe, Bauer, & MacNamara, 2024; Maier & Steinhauser, 2017; Botvinick, Braver, Barch, Carter, & Cohen, 2001).

CDA Dynamics and Postresponse CDAp following the Flanker Response

One key observation is that the CDA began to decrease following flanker task onset and was nearly eliminated by the time participants made their responses, regardless of accuracy. This aligns with recent studies showing that CDA is highly susceptible to interference from concurrently presented stimuli (Kreither, Papaioannou, & Luck, 2022; Hakim et al., 2020, 2021). Following the flanker response, the lateralized waveform was no longer characterized by a canonical negative-going CDA, but instead showed a sustained contralateral positivity, which we refer to here as CDAp. We use this term descriptively to refer to the observed postresponse contralateral positivity, without assuming that it necessarily reflects a canonical CDA with reversed polarity.

Importantly, following the response, CDAp differed between correct and error trials, suggesting that postresponse CDAp may be behaviorally relevant. Notably, despite the decline of the canonical CDA and the emergence of CDAp, participants demonstrated adequate WM performance. This suggests that WM representations may not rely solely on active visual WM storage. Instead, they may be stored in alternative states (Myers, Stokes, & Nobre, 2017; Rose et al., 2016; Nee & Jonides, 2013; Lewis-Peacock, Drysdale, Oberauer, & Postle, 2012), and the CDA may reflect only an early state. For example, items may transition into latent or activity-silent states (Myers et al., 2017; Wolff, Jochim, Akyürek, & Stokes, 2017; Rose et al., 2016; Stokes, 2015), or they may be temporarily held as offline memory representations (Foster, Vogel, & Awh, 2024; Mallett & Lewis-Peacock, 2018; Fukuda & Woodman, 2017). These flexible storage mechanisms are particularly plausible in dual task scenarios, where attention must be reallocated to secondary task-relevant stimuli. In our study, maintaining WM representations across multiple states may have allowed participants to temporarily disengage from active visual WM storage and reallocate resources to process the flanker task.

Adaptive Shielding as a Possible Interpretation of the CDAp

Within this broader framework of state transitions in WM, one possible interpretation is that CDAp reflects a reactive reallocation of attention after flanker interference resolution. One possibility is that individuals who more strongly engage with the flanker task withdraw attentional resources from actively maintained representations to a greater extent, producing a larger CDAp. Importantly, such reactive disengagement does not necessarily imply loss of stored information. Rather, it may reflect temporary offloading or reprioritization of representations into alternative states (Li, Guo, Zhao, & Liu, 2025; Luria & Vogel, 2011).

This interpretation aligns with prior work demonstrating that CDA is sensitive to task goals and attentional priorities. For example, when distractors are expected, their disruptive effect on CDA is reduced, consistent with the notion that a WM gate is closed to protect contents from interference (Kreither et al., 2022; Gresch et al., 2021; Hakim et al., 2021; Lorenc, Mallett, & Lewis-Peacock, 2021; Chatham & Badre, 2015). Likewise, positive-going lateralized shifts have been observed when attention is redirected away from previously maintained items toward task-relevant stimuli (Hakim et al., 2021; Kreither et al., 2022). In addition, CDAp has been observed for distractor locations and interpreted as reflecting sustained suppression of irrelevant information during WM filtering (Feldmann-Wüstefeld & Vogel, 2019).

Importantly, CDAp modulation was not uniform across individuals in the present study. At the group level, CDAp was larger following correct than error responses. Critically, however, the individual-differences analysis revealed that participants who showed relatively greater CDAp modulation following errors than correct responses exhibited smaller ERIAM effects. Thus, the adaptive shielding interpretation is motivated most directly by this brain–behavior relationship.

One possible explanation for this relationship is that CDAp reflects temporary disengagement from active visual WM storage. Participants who disengaged more strongly following errors, possibly by storing items in latent states or offline representations (Foster et al., 2024; Wolff et al., 2017; Stokes, 2015), may have reduced the susceptibility of memory representations to error-related interference. This interpretation is also broadly consistent with the finding that higher WM capacity participants exhibited greater CDAp modulation following responses (partial ρ = 0.45, p = .004, controlling for the ERIAM effect). Although this association does not by itself establish the functional role of CDAp, it further suggests that postresponse CDAp modulation may reflect a regulatory process related to disengagement from active WM storage, rather than a purely passive consequence of attentional withdrawal to the flanker task. Thus, although the present data cannot conclusively establish the underlying mechanism, the relationship between CDAp modulation and the ERIAM effect is consistent with a form of adaptive shielding of WM contents.

One tentative interpretation is that individuals who rely more strongly on this strategy may make more motor errors but experience less WM degradation. Indeed, those with higher flanker error rates tended to show a smaller ERIAM effect (ρ = −0.29, p = .062). Although speculative, this tradeoff suggests that flexible WM strategies reduce interference susceptibility but may impair concurrent task performance. Future studies directly manipulating representational states will be necessary to adjudicate among these possibilities.

Boundary Conditions and Alternative Interpretations of CDAp

Although we interpret CDAp as potentially reflecting adaptive control strategies, alternative explanations warrant consideration. First, prior studies show that CDA amplitude naturally decreases over time (Hakim et al., 2020; Adam et al., 2018; McCollough et al., 2007), suggesting that the observed positivity could reflect a temporal drift. However, in our data, CDA dropped rapidly after the motor response, particularly following correct responses. Importantly, a simple drift would not explain the observed relationship between CDAp and the ERIAM effect, making this account less plausible.

A second possibility is that the observed CDAp reflects the late directing attention positivity—a component associated with orienting or anticipation of an upcoming target at the cued side (McDonald & Green, 2008; Eimer, Velzen, & Driver, 2002; Nobre, Sebestyen, & Miniussi, 2000). Under this interpretation, CDAp could reflect anticipation of the WM probe triggered by the motor response. However, it is unclear how a motor response could act as such a cue well before the WM probe appears.

A third consideration is that the functional significance of CDAp remains uncertain. Consistent with prior work, canonical CDA amplitudes before the flanker task strongly correlated with WM capacity (Figure 4A and B; Luria et al., 2016; Vogel & Machizawa, 2004). After the flanker response, however, this canonical relationship was no longer observed (Figure 4C), suggesting that CDAp may differ from the canonical CDA observed during active maintenance. Moreover, the broader spatial distribution of the postresponse effect makes it difficult to assign CDAp to a single process. Although CDAp modulation persisted after removal of error-related independent components (Supplemental Figure S4), this does not fully resolve its functional origin. Future research is needed to elucidate the neural mechanisms underlying this phenomenon.

Future Research Questions

Our findings raise several avenues for future investigation. First, the current study employed a relatively, short delay between the motor response and the WM probe. However, it is possible that participants reengage visual WM storage after completing the flanker task if the delay period is extended (Clapp, Rubens, & Gazzaley, 2010; McCollough et al., 2007). If so, CDA reactivation might be greater following correct responses than errors. Alternatively, if WM had been reconfigured into activity-silent or alternative storage states (Myers et al., 2017; Lewis-Peacock et al., 2012), CDA might not recover, and condition differences would remain minimal.

Another question concerns the temporal dynamics of error processing (Wessel, 2018). According to this framework, errors trigger automatic inhibition-attentional orienting, followed by strategic adjustments (Choo, Mather, & Wessel, 2023; Guan & Wessel, 2022; King, Korb, von Cramon, & Ullsperger, 2010). An open question is whether the ERIAM effect can be offset by manipulating the timing of posterror WM reactivation. For example, introducing a secondary WM array shortly after the motor error might be ineffective due to lingering error-related inhibition-orienting processes (Jentzsch & Dudschig, 2009; Notebaert et al., 2009). However, allowing recovery time may enable task-set updating and successful WM reinstatement, potentially reducing the ERIAM effect or even improving WM. These avenues can illuminate how cognitive control supports WM under error-related interference.

Conclusion

This study investigated the source of the ERIAM effect (Brochet et al., 2025; Wessel et al., 2022). By tracking CDA, we demonstrated that the ERIAM effect is primarily driven by error commission, rather than general task disengagement or interference from the flanker task. Contrary to our original prediction that errors would produce greater CDA disruption, the postresponse disruption, shown as CDAp, was greater following correct responses than errors. At the same time, greater postresponse CDAp modulation across individuals was linked to smaller ERIAM effects. This pattern might suggest that disengaging from visual WM storage protects memory from interference, reflecting a possible adaptive shielding strategy. However, this interpretation remains speculative. Future research should test this hypothesis using a priori designs aimed at dissociating multistate storage mechanisms and their role in error-related cognitive control.

Acknowledgments

We would like to thank Nathan Chalkley for help with data collection. We also thank the Human Perception and Performance Group (in particular, Andrew Hollingworth, Cathleen Moore, Toby Mordkoff, and Jiefeng Jiang) at The University of Iowa for helpful feedback and suggestions.

Corresponding author: Yoojeong Choo, University of Maryland College Park, College Park, MD, e-mail: yoojchoo@umd.edu.

Data Availability Statement

Data sets as well as related analysis scripts are available at https://osf.io/sjuy7/. Supplemental Material can be accessed on this article's homepage: https://doi.org/10.1162/JOCN.a.2676.

Author Contributions

Yoojeong Choo: Conceptualization; Data curation; Formal analysis; Investigation; Methodology; Software; Visualization; Writing—Original draft; Writing—Review & editing. Kirsten C. S. Adam: Methodology; Supervision; Writing—Review & editing. Jan R. Wessel: Conceptualization; Funding acquisition; Methodology; Project administration; Supervision; Writing—Review & editing.

Declaration of AI and AI-assisted Technologies

The authors declare that the following tool(s) or service(s) were used: chatGPT. AI use for research tasks: not applicable; AI use for manuscript preparation to check grammar.

Funding Information

This work was supported by the National Institutes of Health (https://dx.doi.org/10.13039/100000065; R01NS102201) to J. R. W.

Diversity in Citation Practices

Retrospective analysis of the citations in every article published in this journal from 2010 to 2021 reveals a persistent pattern of gender imbalance: Although the proportions of authorship teams (categorized by estimated gender identification of first author/last author) publishing in the Journal of Cognitive Neuroscience (JoCN) during this period were M(an)/M = .407, W(oman)/M = .32, M/W = .115, and W/W = .159, the comparable proportions for the articles that these authorship teams cited were M/M = .549, W/M = .257, M/W = .109, and W/W = .085 (Postle and Fulvio, JoCN, 34:1, pp. 1–3). Consequently, JoCN encourages all authors to consider gender balance explicitly when selecting which articles to cite and gives them the opportunity to report their article's gender citation balance. The authors of this paper report its proportions of citations by gender category to be: M/M = .722; W/M = .167; M/W = .069; W/W = .042.