Maintenance suppression enhances subsequent associative learning

Significance

Suppressing irrelevant information in working memory enhances the long-term formation of subsequent memories. Through behavioral and neuroimaging experiments, we demonstrate that intentionally suppressing, just-learned associations facilitate the encoding of new ones, supported by enhanced neural fidelity in the hippocampal subfield CA1. Moreover, memory retrieval is facilitated by a neural bias away from previously suppressed contents. These findings highlight the adaptive role of active forgetting in optimizing learning and memory.

Abstract

Removing irrelevant information from working memory (WM) can free cognitive resources and reduce interference with current task goals. Beyond these immediate benefits, removal may also support long-term memory processes. We tested this hypothesis using an associative memory paradigm with directed forgetting instructions and characterized the underlying neural mechanisms. In complementary behavioral (N = 22) and fMRI (N = 17) experiments, participants completed an ABC associative memory task using pictures of objects, faces, and scenes. Each A item was sequentially paired with B and C items. Our key manipulation targeted the B items: following B encoding, participants were instructed either to maintain or suppress B. We hypothesized that suppression would impair AB memory but enhance AC memory, and this is precisely what we found. Suppression cues elicited distributed activations across frontoparietal control regions, as expected. Critically, multivoxel pattern analyses revealed enhanced encoding fidelity for C items in hippocampal subregion CA1 following suppression compared to maintenance. During subsequent associative memory tests cued by A items, reactivation of suppressed B items was significantly reduced relative to C items, reflecting diminished competition from B items during retrieval. Together, these findings demonstrate that suppression of unwanted information in WM benefits the formation of new memories by enhancing their encoding fidelity and biasing subsequent memory retrieval away from the suppressed associations.

Interference from related memories is a significant cause of undesired forgetting in human memory (1–3). Forgetting competing memories has been hypothesized to serve an adaptive function, enhancing the retention of more important target memories (4–7). For example, when an individual changes their email password, breaking the association between the email account and the old password might help prevent interference and potentially facilitate the formation of a stronger association with the new password. Active forgetting of unwanted information can occur both during encoding (e.g., maintenance suppression; 8–10) and during retrieval (e.g., retrieval suppression; 5, 11–13) of a memory episode. While both forms of suppression have been shown to reduce the accessibility of memories targeted for suppression, it remains unclear how these processes also facilitate the formation and retention of new long-term memories (e.g., a new password).

Memory suppression during the encoding is characterized by the engagement of multiple frontoparietal control regions (8, 14). Working memory (WM) operation cues that instruct participants to suppress an encoded item (maintenance suppression) elicit greater activation in the dorsal anterior cingulate cortex (dACC) and dorsolateral prefrontal cortex (dlPFC) compared to cues that direct participants to actively maintain an item (8, 9). Similarly, cues instructing participants to forget a just-encoded item (item-method directed forgetting) have been shown to increase activation in several frontoparietal regions, including the DLPFC, insula, right superior frontal gyrus, and right inferior frontal gyrus compared to a visually matched maintain condition (14–18). This consistent recruitment of frontoparietal control regions, alongside the eventual forgetting observed after suppression, suggests that to-be-suppressed memories are not simply ignored. Instead, they are actively targeted and purged from memory through top–down control mechanisms.

Active suppression of a memory may facilitate the encoding and retention of other memories. This hypothesis has been explored using list-method directed forgetting paradigms, where suppressing a previously studied list of words not only reduces recall accuracy for that list but also enhances learning of a subsequent list (19–22). It has been proposed that such benefits of directed forgetting on subsequent learning may arise from the engagement of deeper encoding strategies by participants (21, 22). Related, given that encoding relies on a capacity-limited WM system, removing irrelevant content may optimize the allocation of encoding resources, improving the fidelity of newly formed representations (9, 23–25).

Beyond that, suppression may also reduce interference by weakening the representations of to-be-suppressed information. Typically, previously encoded information lingers in WM if not actively removed (26). While this feature benefits memory stability in continuous experiences (e.g., serial dependency; 27–29), lingering representations can cause interference when contexts shift, and new memories need to be formed or prioritized (26, 30–34). Suppression may function to purge these residual traces from WM, thereby mitigating interference and enabling more effective encoding of new information. These two mechanisms, freeing encoding resources and reducing proactive interference, are not mutually exclusive and may jointly support memory formation following suppression.

To our knowledge, how the brain facilitates memory following suppression has not been previously examined. It remains unclear which brain regions involved in memory encoding may be modulated by a preceding act of suppression. However, the memory encoding literature points to two strong candidates, particularly in the context of attention-driven modulation effects during encoding. The first is the sensory cortex. Selective attention has been shown to bias sensory cortices toward representing attended information, which in turn enhances hippocampal-supported memory (35–37). In line with this framework, recent evidence suggests that encoding fidelity, measured using representational similarity analysis (RSA) in the ventral temporal cortex (VTC), is enhanced immediately following suppression. This finding implies that suppression may enhance the encoding of subsequently presented information in sensory regions and thus strengthen the formation of new memory (9).

The other candidate is the hippocampus. Although prior studies have reported mixed findings regarding the modulation of overall hippocampal activity by attention (38–41), more recent work using RSA suggests that memory representations in hippocampal subregions can be stabilized by directing attention to relevant content (42, 43). The specific contributions of hippocampal subregions to memory encoding have been actively investigated. One line of work shows that the hippocampal input structures, CA2/3 and the dentate gyrus (collectively CA23/DG), are selectively active during encoding and that their activation predicts subsequent memory success (44–46). Attention enhances representational stability primarily in CA23/DG, with less consistent effects in CA1 (42, 43).

Parallel research suggests that CA1 plays a distinct role during encoding, particularly in comparing or linking incoming sensory information to existing memories (47, 48). CA1 receives input both from CA3 that conveys memory signals and directly from the entorhinal cortex, which conveys sensory signals (47, 49, 50). During encoding, sensory representations are prioritized via enhanced connectivity between CA1 and entorhinal cortex (51–53), particularly when sensory input mismatches mnemonic predictions, potentially biasing hippocampal states toward enhanced encoding (54). In addition, CA1 is sensitive to stimulus novelty and shows increased activation when novel information needs to be encoded (55, 56). Attention instructions that emphasize the distinctiveness of visual stimuli have also been shown to increase CA1 activity and strengthen memory formation (57). Finally, CA1 has been uniquely implicated in memory integration. Reactivation of prior memories in CA1 during new learning has been associated with the integration of new sensory inputs into existing memory (58–60). Together, a rich body of literature suggests that hippocampal subregions are actively engaged in associative memory formation and can be directly modulated by cognitive control mechanisms.

Here, we sought to characterize the neural mechanisms that support potential memory facilitation following suppression in WM during encoding. To test this, we conducted a human fMRI experiment and a complementary behavioral experiment using an ABC associative learning task. On each trial, participants were instructed to associate an A item with subsequently presented B and C items. Our critical manipulation involved the B item: Participants were instructed either to maintain B or to suppress B. Each trial concluded with a C item that participants were asked to associate with the preceding A item. We hypothesized that suppression would impair AB memory formation but enhance the subsequent encoding of the AC association. We then examined how suppression modulates memory encoding processes in two relevant brain systems: the VTC, which represents sensory information, and the medial temporal lobe (MTL), which supports the formation of new memories. We focused in particular on patterns of neural activity during suppression of B items, during learning of C items, and during subsequent cued retrieval of these items. Motivated by prior work demonstrating distinct roles of hippocampal subregions in memory encoding, we leveraged the high spatial resolution of our imaging protocol to interrogate specific hippocampal subfields, including CA1 and CA23/DG, to examine their contributions to associative memory formation following the suppression of information from WM.

Results

After each run of 12 learning trials (4 trials per condition), we tested participants’ associative memory for the just-learned pairs. In the testing phase, a cue image (A item) was presented first, followed by four probe images (Fig. 1B). Participants were asked to identify which of the four probes had been associated with the cue image. Each cue was tested only once. This means that in the maintain and suppress conditions where the cue was associated with both B and C images, we tested either the AB or the AC pair, but never both. AB and AC pairs had an equal probability of being tested. In the fMRI study, the response window was fixed at 4 s to align with the TR, while in the complementary behavioral study, the response window ended when participants made a response.

Fig. 1.

Task illustrations and behavioral results. (A) Participants were instructed to form AB and AC associations. The key manipulation pertains to B items. In the suppress B condition, participants suppressed the just-presented image B, whereas in the maintain B condition, participants tried to form AB associations. In the no B condition, no image B was presented. Our analyses pertain to three key windows: operation phase, encode c phase, and the test phase. (B) We hypothesized that the suppression of B items would weaken AB associations but strengthen AC associations. (C) Memory results from the main fMRI experiment: memory accuracy for AB pairs was impaired in the suppress B condition compared to the maintain B condition. Memory accuracy for AC pairs in the maintain condition was impaired compared to the baseline condition, but no memory cost was found in the suppress condition. (D) Memory results from the complementary behavioral experiment: accuracy for AB pairs was impaired following suppressions compared to the maintain B condition. Memory accuracy for AC pairs in the maintain condition was impaired compared to both the baseline and the suppress condition, but no memory cost was found in the suppress condition. * indicates corrected P < 0.05; ** indicates corrected P < 0.01; *** indicates corrected P < 0.001; (*) indicates uncorrected P < 0.05 but corrected P > 0.05.

Suppression Impaired AB Memory but Facilitated AC Memory.

Memory performance for AB pairs was impaired in the suppress condition (Fig. 1C and SI Appendix, Fig. S1A). The mean memory accuracy in the maintain condition was significantly higher than in the suppress condition, t(16) = 6.04, P < 0.001, d = 1.71, BF10 = 1359.60. Additionally, the mean reaction time was significantly faster in the maintain condition compared to the suppress condition, t(16) = −5.80, P < 0.001, d = 1.94, BF10 = 908.05.

Consistent with our hypothesis, associative memory performance for AC pairs benefited from the suppression of B images (Fig. 1C and SI Appendix, Fig. S1A). The mean memory accuracy for AC pairs was not different between conditions, F(2,32) = 2.39, P = 0.108, ${\eta }_p^2$ = 0.13. The accuracy for AC pairs in the maintain condition was slightly impaired relative to the baseline condition [t(16) = −2.17, P = 0.045, corrected P = 0.135, d = 0.36, BF10 = 1.60]. No significant difference was observed, however, between the suppress condition and the no B baseline [t(16) = 0.26, P = 0.796, corrected P = 0.796, BF10 = 0.26], nor between the suppress condition and the maintain condition [t(16) = 1.67, P = 0.114, corrected P = 0.172, BF10 = 0.79]. Memory benefits following suppression of B items were more reliable in reaction time. Mean reaction time significantly differed between conditions F(2,32) = 8.05, P = 0.001, ${\eta }_p^2$ = 0.33. The mean reaction time for AC pairs was significantly slower in the maintain condition compared to the baseline [t(16) = 3.73, P = 0.002, corrected P = 0.005, d = 0.74, BF10 = 22.55] or the suppress condition [t(16) = 2.41, P = 0.028, corrected P = 0.042, d = 0.58, BF10 = 2.32]. No significant difference between the suppress and baseline conditions was found, t(16) = −1.33, P = 0.203, corrected P = 0.203, BF10 = 0.53.

Those results were largely consistent with the complementary behavioral experiment (Fig. 1D and SI Appendix, Fig. S1B). Memory performance for AB pairs was impaired in the suppress condition. The mean memory accuracy for AB pairs in the maintain condition was higher than in the suppress condition, t(21) = 3.98, P < 0.001, d = 0.57, BF10 = 49.89. The mean reaction time for AB pairs was faster in the maintain than the suppress condition, t(21) = −5.62, P < 0.001, d = 0.80, BF10 = 1616.85.

Critically, associative memory performance for AC pairs benefited from the suppression of B images (Fig. 1D and SI Appendix, Fig. S1B). The mean memory accuracy for AC pairs differed between conditions, F(2,42) = 11.19, P < 0.001, ${\eta }_p^2$ = 0.08. The maintain condition showed impaired AC performance relative to both the baseline and the suppress condition [maintain vs baseline: t(21) = −4.23, P < 0.001, corrected P = 0.001, d = 0.68, BF10 = 84.47; maintain vs suppress: t(21) = −3.12, P = 0.005, corrected P = 0.008, d = 0.46, BF10 = 8.60]. No significant difference was observed, however, between the suppress condition and the baseline, t(16) = 1.51, P = 0.147, corrected P = 0.147, BF10 = 0.60. Mean reaction time also significantly differed between conditions, F(2,42) = 5.72, P = 0.006, ${\eta }_p^2$ = 0.04. The mean reaction time for AC pairs was significantly slower in the maintain condition compared to the baseline [t(21) = 3.00, P = 0.007, corrected P = 0.020, d = 0.49, BF10 = 6.90], but not the suppress condition [t(21) = 1.07, P = 0.298, corrected P = 0.298, BF10 = 0.37]. The mean reaction time for AC pairs in the suppress condition was slower than the baseline condition, t(21) = 2.36, P = 0.028, corrected P = 0.042, d = 0.40, BF10 = 2.15.

The memory benefits of suppressing B items on AC performance manifested primarily as faster reaction times in the fMRI study, whereas they appeared mainly as accuracy benefits in the complementary behavioral study. This difference is likely due to a design difference. In the fMRI study, participants were required to respond within a 4 s window (mean RTs between 1.60 and 2.47 s in all conditions), likely encouraging them to prioritize response time over accuracy, whereas in the behavioral study, no time limits were imposed, allowing participants to prioritize accuracy over response time (mean RTs > 4.80 s in all conditions).

Suppression Engaged Distributed Frontal–Parietal Regions.

We identified brain regions that support inhibitory control during maintenance suppression. Replicating prior studies (8–10), univariate contrasts between the suppress and maintain conditions revealed that suppression evoked conjoint activations in multiple frontal–parietal regions implicated in inhibitory control (Fig. 2A, see SI Appendix, Table S1 for specific MNI coordinates), suggesting that suppression likely involved active control processes rather than a passive release of information. These regions included the dlPFC, ventrolateral prefrontal cortex (vlPFC), insula, dACC, temporoparietal junction (TPJ), and posterior cingulate cortex (PCC). Consistent with prior findings, right PFC regions were more closely related to the suppression operation compared to the left hemisphere (14).

Fig. 2.

Suppression engaged distributed frontal–parietal regions. (A) Univariate comparisons between the suppression and maintain conditions revealed that suppression cues elicited stronger activations in multiple frontal–parietal regions, including the dlPFC, vlPFC, dACC, insula, TPJ, and PCC. (B) WM operations (suppression vs. maintenance) can be reliably decoded from whole-brain activation patterns. *** indicates corrected P < 0.001.

We then trained an operation classifier on whole-brain data for each participant to differentiate multivariate patterns associated with the maintain and suppress conditions. Classification performance was assessed using the area under the receiver operating characteristic curve (AUC), which was significantly higher than the chance level of 0.5 for both operations (Fig. 2B, M = 0.70, SD = 0.13, t(16) = 6.33, P < 0.001, d = 1.54, BF10 = 4419.26), which replicates prior observations (9, 10).

Enhanced Encoding Fidelity of C Items in CA1 Following Maintenance Suppression.

The primary goal of this study was to investigate whether maintenance suppression benefits subsequent learning of new associations. The visual presentations and task requirements were identical during the learning of C items across all conditions. The difference was whether participants were asked to maintain the preceding B item or to suppress it, or if there was no B item presented (baseline). Encoding C items following B items could be impaired due to lingering representations of B items or due to reduced availability of WM resources. Consequently, the neural representation of C items that follow B items may be less reliable than when no B items were present. Thus, we hypothesized that suppressing B items would lead to more reliable encoding of C items, compared to when B items were actively maintained. To test this, we assessed the encoding fidelity of C items by calculating the average similarity between the neural pattern of each C item and other C items from the same category, compared to C items from the opposite category (Fig. 3A). This RSA-based approach is commonly used in small ROIs, such as hippocampal subregions, where decoding is less feasible due to limited voxel counts (61–63). Notably, we replicated our findings using both RSA and MVPA decoding analyses in these regions. We focused on three brain regions of interest that are involved in memory encoding and learning: CA1, CA23/DG, and VTC.

Fig. 3.

Enhanced encoding fidelity of C items in CA1 following maintenance suppression. (A) Illustration of the RSA approach. Category RSA was calculated for CA1 and CA23/DG as the difference between the correlation of images within the same category and the correlation of images from different categories. (B) Representational similarity analyses showed reliable encoding fidelity of C items in CA1 in the no B and suppress B condition, but not in the maintain B condition. Critically, suppression led to enhanced encoding fidelity of C items compared to the maintain condition. In CA23/DG, no reliable encoding fidelity was observed across conditions. (C) MVPA decoding approaches showed reliable decoding of C items in CA1 in the no B and suppress B condition, but not in the maintain B condition. Critically, suppression led to enhanced decoding evidence of C items compared to the maintain condition. In CA23/DG, no reliable decoding of C items was observed across conditions. Horizontal lines indicate the theoretical chance levels: 0 for category RSA, and 0.5 for MVPA decoding evidence. * indicates corrected P < 0.05; ** indicates corrected P < 0.01; (*) indicates uncorrected P < 0.05 but corrected P > 0.05.

Encoding fidelity varied across conditions only in CA1 (Fig. 3B), F(2, 28) = 4.27, P = 0.024, ${\eta }_p^2$ = 0.23. The encoding fidelity in the baseline and suppress conditions was reliably above 0 [baseline, t(14) = 2.98, P = 0.005, corrected P = 0.015, d = 0.77, BF10 = 11.13; suppress, t(14) = 2.14, P = 0.025, corrected P = 0.038, d = 0.55, BF10 = 3.07; maintain, t(14) = −0.09, P = 0.536, corrected P = 0.536, BF10 = 0.53], and significantly greater than the maintain condition [baseline—maintain, t(14) = 2.50, P = 0.025, corrected P = 0.064, d = 0.84, BF10 = 2.61; suppress—maintain, t(14) = 2.23, P = 0.043, corrected P = 0.064, d = 0.52, BF10 = 1.74].

In CA23/DG, encoding fidelity was not reliably positive across conditions (F(2,28) = 0.03, P = 0.970, ${\eta }_p^2$ = 0.00; baseline, t(14) = 1.23, P = 0.120, corrected P = 0.120, BF10 = 0.99; suppress, t(14) = 1.60, P = 0.066, corrected P = 0.120, BF10 = 1.48; maintain, t(14) = 1.33, P = 0.102, corrected P = 0.120, BF10 = 1.10). In the VTC (SI Appendix, Fig. S2A), encoding fidelity was reliably positive across conditions [baseline, t(16) = 6.00, P < 0.001, corrected P < 0.001, d = 1.45, BF10 = 2535.12; suppress, t(16) = 6.16, P < 0.001, corrected P < 0.001, d = 1.49, BF10 = 3347.67; maintain, t(16) = 5.46, P < 0.001, corrected P < 0.001, d = 1.32, BF10 = 1003.63], but no differences between conditions were found, F(2,32) = 1.79, P = 0.183, ${\eta }_p^2$ = 0.10.

We replicated the encoding fidelity results with MVPA approaches (Fig. 3C). Classifiers were trained on the brain scan when the C item was presented (shifted 4 s for hemodynamic lag). We used a leave-one-run-out cross-validation approach to analyze the data. Classifiers were trained on all but one run and tested on the remaining run of data. Decoding evidence of C items varied across conditions only in CA1 (Fig. 3C), F(2,28) = 5.74, P = 0.008, ${\eta }_p^2$ = 0.29. The classifier evidence of C items in the baseline and suppress conditions did not differ but was reliably above the theoretical chance level of 0.5 [baseline, t(14) = 3.32, P = 0.003, corrected P = 0.008, d = 0.86, BF10 = 19.75; suppress, t(14) = 2.13, P = 0.026, corrected P = 0.039, d = 0.55, BF10 = 3.00; maintain, t(14) = −2.21, P = 0.978, corrected P = 0.978, BF10 = 0.30], and significantly greater than the maintain condition [baseline—maintain, t(14) = 4.05, P = 0.001, corrected P = 0.004, d = 1.43, BF10 = 33.01; suppress—maintain, t(14) = 2.68, P = 0.018, corrected P = 0.038, d = 1.09, BF10 = 3.44; baseline—suppress, t(14) = 0.11, P = 0.916, corrected P = 0.916, BF10 = 0.26]. The observed reduction in decoding evidence of C items in CA1 following the maintenance of B items suggests a neural cost associated with the prior encoding of Bs. In contrast, suppression of B items mitigated this cost, leading to more reliable encoding patterns for C items in CA1.

In CA23/DG, decoding evidence of C items was not reliably above the chance level of 0.5 across conditions [F(2,28) = 2.41, P = 0.108, ${\eta }_p^2$ = 0.15; baseline, t(14) = 0.82, P = 0.212, corrected P = 0.318, BF10 = 0.70; suppress, t(14) = 1.11, P = 0.142, corrected P = 0.318, BF10 = 0.89; maintain, t(14) = −1.71, P = 0.946, corrected P = 0.946, BF10 = 0.58]. In the VTC (SI Appendix, Fig. S2B), classifier-based encoding fidelity was reliable across conditions [baseline, t(16) = 5.96, P < 0.001, corrected P < 0.001, d = 1.45, BF10 = 2535.12; suppress, t(16) = 7.42, P < 0.001, corrected P < 0.001, d = 1.80, BF10 = 2549.0; maintain, t(16) = 5.92, P < 0.001, corrected P < 0.001, d = 1.44, BF10 = 2256.36], but no differences between conditions were found, F(2,32) = 2.72, P = 0.081, ${\eta }_p^2$ = 0.15.

We identified that suppression led to distinct activation/deactivation patterns during the operation phase following the encoding of B items and enhanced CA1 encoding fidelity in the subsequent learning phase for C items. Next, we addressed the functional significance of these two findings. We tested whether these neural consequences of suppression were associated with memory performance, focusing on performance differences between the suppress and maintain conditions for both AB and AC pairs. Behaviorally, suppression resulted in more forgetting of AB pairs and a slight benefit for AC memory compared to the maintain condition. Higher classifier AUC values for suppression and greater CA1 encoding fidelity scores (reflecting stabilization of the memory trace) may reflect participants’ engagement in the suppression operation and the enhanced encoding of subsequent C items. Based on this, we predicted that both neural indices would be associated with forgetting of AB pairs and facilitation of AC memory formation. We conducted between-subject linear models to test those potential associations (see SI Appendix, Fig. S5 for detailed descriptions of those linear models). Both suppression-related neural patterns and the subsequent CA1 stabilization patterns predicted the forgetting of AB pairs independently (SI Appendix, Fig. S5A, Operation AUC: $\beta = 0.12$, P = 0.021, BF10 = 3.48; SI Appendix, Fig. S5B, CA1 RSA: $\beta = 0.12$, P = 0.023, BF10 = 3.21). However, only CA1 encoding fidelity positively predicted AC facilitation (SI Appendix, Fig. S5D, $\beta = 0.10$, P = 0.014, BF10 = 2.25). Given our limited sample size, interpretations of the brain–behavior correlations should be made with caution.

Suppression Biased Subsequent Retrieval Away from Suppressed Items.

The preceding analyses indicated that suppression stabilized CA1 activity patterns during the encoding of new C items. We next examined how these encoding differences influenced memory retrieval. Prior research has linked memory retrieval to the reinstatement of sensory cortical patterns (64, 65). Therefore, to independently track categorical evidence for B items and C items, we first trained a category classifier using fMRI data from an independent perceptual localizer task, capturing cortical representations of different image categories during sensory processing. We then applied this trained classifier to track the evolution of categorical representations throughout the retrieval and decision components of the test phase.

When testing image pairs initially learned in the maintain condition, there was a categorical bias toward the reactivation of B category information following the onset of probe images. Specifically, the categorical evidence of B items was stronger than that of C items [Fig. 4B, 10 s, t(16) = 2.72, P = 0.015, corrected P = 0.096, d = 0.68, BF10 = 3.85; 12 s, t(16) = 2.61, P = 0.019, corrected P = 0.096, d = 0.63, BF10 = 3.16], indicating that retrieval was biased to B items for the maintain B condition. Intriguingly, when testing image pairs learned in the suppress condition, a reversed categorical bias was observed, favoring C items. The categorical evidence of C items was stronger than that of B items [10 s, t(16) = 3.03, P = 0.008, corrected P = 0.047, d = 0.73, BF10 = 6.46; 12 s, t(16) = 2.95, P = 0.009, corrected P = 0.047, d = 0.71, BF10 = 5.63]. The bias toward C items was stronger in the suppress B condition compared to the maintain B condition [10 s, t(16) = 3.84, P = 0.001, corrected P = 0.012, d = 1.40, BF10 = 27.74; 12 s, t(16) = 3.58, P = 0.002, corrected P = 0.012, d = 1.35, BF10 = 17.33].

Fig. 4.

Suppression led to a memory bias toward C items in the VTC and precuneus during memory retrieval. (A) Illustration of memory retrieval and decision making in the maintain, suppress, and no B condition. (B) Machine learning classifiers were trained on perceptual localizer data and were used to track category evidence during memory tests. Category biases were computed as the differences between category C evidence and category B evidence. In the maintain B condition, categorical evidence of B items was stronger than that of C items, indicating memory biases toward Bs. Conversely, in the suppress B condition, categorical evidence of C items was stronger than that of B items. (C) Comparisons of category C evidence at its peak time showed the strongest evidence in the baseline condition, followed by the suppress condition, and the weakest in the maintain condition. (D) Whole-brain searchlight analysis showed that to-be-retrieved categorical information was represented in PCUN during the retrieval phase of the No B condition. The classifier trained on PCUN activities during the retrieval phase of the No B condition revealed retrieval biases toward C items in the suppress B condition, whereas no reliable retrieval biases were found in the maintain B condition. The purple lines indicate the moments when category C biases reached their peak. * indicates corrected P < 0.05; ** indicates corrected P < 0.01; *** indicates corrected P < 0.001; (*) indicates uncorrected P < 0.05 but corrected P > 0.05. (**) indicates uncorrected P < 0.01 but corrected P > 0.05. Green or orange * indicates statistical significance when comparing the corresponding maintain B or suppress B condition to the theoretical chance level. Black * indicates statistical significance when comparing between the maintain B and suppress B conditions.

In addition, the categorical evidence of C items at its peak time of the baseline condition (12 s, for the full time course, see SI Appendix, Fig. S7) during retrieval differed between conditions, F(2,32) = 30.05, P < 0.001, ${\eta }_p^2$ = 0.65. The categorical evidence of C items was strongest in the baseline condition, followed by the suppress condition, and weakest in the maintain condition [suppress—maintain, t(16) = 3.53, P = 0.002, corrected P = 0.005, d = 1.07, BF10 = 15.77; baseline—maintain, t(16) = 6.83, P < 0.001, corrected, P < 0.001, d = 2.09, BF10 = 5003.05; baseline—suppress, t(16) = 4.90, P < 0.001, corrected P < 0.001, d = 1.02, BF10 = 188.49]. These findings indicate that suppression enhanced the retrieval of the more recently learned C items compared to the maintain B condition.

The categorical bias toward C items in the suppress condition emerged after the onset of the probe images (unshifted for the hemodynamic lag), suggesting that this bias likely reflected both a retrieval bias toward C items and an attentional bias during the evaluation of probe images that matched C’s category. To more specifically isolate retrieval-related biases, as opposed to attentional biases driven by probe processing, we conducted an additional analysis focused on the retrieval phase of the memory test.

Beyond the reactivation of sensory representations in sensory regions, memory retrieval has also been linked to widely distributed, retrieval-unique representations in frontal and parietal cortices (65, 66). These findings suggest that retrieval not only reactivates prior sensory traces but may also transform them into distinct formats, encoded in different brain regions. To assess these retrieval-unique representations, we examined activity during the retrieval phase (4 to 8 s following cue onset and prior to probe onset) of the baseline condition. We performed a whole-brain searchlight analysis to identify clusters of voxels that carried categorical information (face vs. scene) about the to-be-retrieved items. The largest cluster of voxels was found in the precuneus (Fig. 4D). We then trained a two-category classifier (face vs. scene) based on this retrieval-related activity in the precuneus from the baseline condition. The decoder model was trained on the baseline condition only, so it was independent of the testing data from the maintain and suppress conditions. We then applied this classifier to the other conditions to capture the evolution of retrieval-related categorical representations. Consistent with VTC biases that were observed, the categorical evidence of C items in the suppress condition was above chance [Fig. 4D, 10 s, t(16) = 2.59, P = 0.010, corrected P = 0.099, d = 0.63, BF10 = 6.15], and stronger than in the maintain condition 2 s after the onset of probe images (unshifted for hemodynamic lag) [10 s, t(16) = 2.98, P = 0.009, corrected P = 0.088, d = 0.98, BF10 = 6.00]. The peak time of C item representation in the precuneus was earlier than that in the VTC but was delayed relative to the baseline condition. This analysis suggests that the categorical biases between maintain B and suppress B conditions was at least partially driven by retrieval activity and not solely driven by attentional biases during the processing of the probe items.

We then examined and found a potential functional link between retrieval-related neural biases and memory outcomes (see SI Appendix, Fig. S6 for detailed descriptions of those between-subject linear models). VTC bias significantly predicted the forgetting of AB pairs in the suppress condition (SI Appendix, Fig. S6A, Adjusted $R^2$ = 0.32, F = 8.53, P = 0.011, BF10 = 3.90), and precuneus bias significantly predicted facilitation of AC performance in the suppress condition relative to the maintain condition (SI Appendix, Fig. S6D, Adjusted $R^2$ = 0.24, F = 5.95, P = 0.028, BF10 = 1.12). Given our limited sample size, interpretations of the brain–behavior correlations should be made with caution.

Discussion

Theoretical perspectives have suggested that the adaptive forgetting of unwanted memories can benefit the retention of other memories (4–7). Although previous studies have investigated this idea using behavioral list-method directed forgetting paradigms (19–22), the neural mechanisms underlying suppression-induced memory facilitation remain surprisingly poorly understood. While some neural evidence exists for facilitation after retrieval suppression (7), the mechanisms underlying maintenance suppression remain underexplored. Here, we used pattern-based fMRI measures of hippocampal and sensory memory fidelities in humans to address this gap. Consistent with prior work, we first demonstrated that maintenance suppression relies on the engagement of frontal–parietal control regions. Importantly, our findings reveal that suppressing a previously encoded association (AB) facilitated the encoding of new associations (AC), as reflected by increased encoding fidelity of C items in the hippocampal CA1 subregion. Furthermore, during cued retrieval, we observed neural evidence of retrieval facilitation when the B item of an ABC triplet had been suppressed. Specifically, stronger categorical reactivation of C items compared to B items was observed in both the VTC and the precuneus. In contrast, when the B items had been maintained, there was a stronger reactivation of B items during the retrieval. These findings provide robust support for the associative memory facilitation hypothesis of maintenance suppression. They elucidate the neural mechanisms by which maintenance suppression enhances the encoding and retrieval of new, goal-relevant information.

In the suppress condition, participants were instructed to actively “push the item memory down every time it comes to mind” (full instructions provided in Methods) to encourage engagement in the suppression process. An alternative hypothesis is that participants either aborted encoding after seeing the suppress cue or delayed encoding until the cue appeared, deciding whether or not to encode image B based on the cue. Evidence supporting this aborted encoding hypothesis would be if the maintain cue elicited unique neural patterns associated with maintenance, while such patterns were diminished or absent in the suppress condition. However, we found that the suppress cue elicited even stronger activation in multiple frontoparietal regions, suggesting that suppression was not due to passive disengagement but was instead supported by active inhibitory control processes requiring greater engagement of cognitive control mechanisms. Our finding that maintenance suppression involved contributions from multiple frontal–parietal regions is also consistent with prior research. Activations in the dACC, DLPFC, insula, SFG, and IFG have been consistently observed in both item-method directed forgetting paradigms (14–18) and maintenance suppression paradigms (8, 9), indicating a general inhibitory control mechanism contributing to the suppression of both long-term memory encoding and maintenance in WM. Moreover, recent cross-domain comparisons have converged on the view that inhibitory control across modalities arises from common, domain-general neural mechanisms (67, 68). For example, similar frontal activation patterns have been observed in studies where participants were instructed to stop the retrieval of certain memories or to stop the execution of actions (68). It remains unclear whether retrieval suppression and maintenance suppression are supported by overlapping mechanisms. Future research should investigate this possibility, exploring both the commonalities and distinctions between different forms of memory suppression and their relationship to inhibitory control.

While the mechanisms underlying maintenance suppression have been well characterized (8–10), its effects on the formation of new memories remain less well understood. We hypothesized that maintenance suppression might free up WM resources and/or reduce proactive interference, thereby facilitating the encoding of new information. We specifically tested two neural candidates during the encoding phase following suppression: the VTC and the hippocampus. In trials where ABC items were sequentially presented, neural representations of C items were detectable only in CA1 following suppression, but not in the maintain condition, suggesting that suppression enhanced the encoding of C items in CA1. In contrast, representations of C items were reliably detectable in the VTC across conditions but were not modulated by the prior memory operation (maintain vs. suppress), suggesting that the VTC may primarily reflect veridical incoming sensory information.

The enhanced encoding observed in CA1 could result from increased neural resources or capacity following suppression. In the maintain condition, participants were instructed to encode two items (A and B) before encoding the C item, whereas in the suppress condition, they were again instructed to encode A and B items but then to suppress B before encoding C. The reduced memory retention demands in the suppress condition could have increased the availability of neural resources for encoding C items. Alternatively, the enhanced CA1 representations could reflect reduced representational competition between B and C items. Prior research has shown that memory competition often leads to ambiguous neural patterns and diminished fidelity (69). Thus, the finding that encoding fidelity for C items was enhanced in the suppress condition compared to the maintain condition suggests that suppression may help safeguard memory encoding against proactive interference.

The hippocampal subregion CA1 plays an important role in linking sensory experiences with existing memories. Prior studies have shown that CA1 is activated both when novel experiences are encountered and when mnemonic predictions are violated (53–56). In both scenarios, the increased need for encoding may bias the hippocampus toward an encoding state, potentially by prioritizing sensory inputs conveyed via CA1–entorhinal connections (54). CA1 also facilitates the integration of sensory information with prior memories by reactivating existing memory traces and enhancing the encoding of related sensory information (58, 59). Consistent with the role of CA1 in signaling the need for encoding, the strengthening of C item representations following suppression in our study could reflect enhanced encoding processes. Alternatively, given the sequential presentation of A and C items, CA1 might also have been recruited to support the binding of sensory representations of Cs to memory representations of As. Future studies should investigate whether CA1 memory representations are similarly enhanced following suppression in nonassociative learning tasks, to determine whether CA1 broadly supports enhanced encoding or plays a more specific role in associative memory formation. Prior research has also shown that maintenance suppression can enhance encoding fidelity in the VTC for subsequent items during WM tasks (9). The discrepancy between this and our current results may reflect the flexibility of inhibitory mechanisms to modulate representations in task-relevant regions. Whereas VTC representations are critical for WM maintenance, CA1 representations are key for forming associative memories. This flexibility suggests that the target of suppression might adapt depending on task demands.

We examined the long-term neural consequences of maintenance suppression by tracking the evolution of categorical evidence during the retrieval and response phases of this study. In the maintain condition, stronger categorical evidence for B items compared to C items was observed in the VTC, suggesting that representations of B items dominated during the retrieval and response phases in memory tests. This is not maladaptive per se but indicates that the A items were a stronger cue for the more temporally proximal B items than for the non–temporally adjacent C items. With similar designs, prior studies using MVPA of EEG and fMRI data have shown that proactive interference was related to early reinstatement of the competitor context (70), and less reliable target reactivation (69, 71). Importantly, the reversed pattern—with stronger categorical evidence for C items than B items—was found in the suppress condition. This provides critical neural evidence that maintenance suppression during the encoding phase biases the competition between memories during the retrieval phase, facilitating the retrieval of nonsuppressed memories over suppressed memories. However, it is worth noting that the difference in categorical evidence between items B and C only emerged after the onset of probe images (unshifted for hemodynamic lag). This suggests that during the test, participants may have shifted their attentional focus to the category of C items instead of B items for cues in which they had previously suppressed the B item.

To capture a more accurate retrieval-related signal, we focused on the retrieval phase before probe onset in the baseline condition and performed a whole-brain searchlight analysis to locate retrieval-related representations in the brain. This analysis revealed reliable representations for to-be-retrieved C item categories in the precuneus. The precuneus has been implicated in memory retrieval, particularly for autobiographical memories (66, 72, 73). By training classifiers with precuneus signals in the baseline condition, we found a neural bias toward representing C item category during retrieval only in the suppress condition but not in the maintain condition. This indicates that maintenance suppression during the encoding phase produced biases in memory retrieval away from the suppressed item during the subsequent retrieval phase. However, this bias also occurred shortly after probe onset, which might indicate that the retrieval process in the suppress condition was delayed compared to the baseline condition (65). One critical assumption of this analysis is that the retrieved representations in the baseline condition should be generalizable to retrieved representations in other conditions. This assumption is not guaranteed to be true, given that memory representations can be reformatted, such as through integration or differentiation, to reduce interference (66). Future research should examine whether memories encoded with interfering information undergo transformation compared to those encoded without interference.

Memory interference, particularly between overlapping events, is a primary contributor to forgetting in the human brain. Previous research has highlighted integration and differentiation as potential strategies for mitigating such interference (62, 74, 75). When overlapping information must be maintained in the memory system, representations of these events can be adaptively modulated through hippocampal differentiation or integration to reduce interference (66). However, in situations where certain information becomes irrelevant or no longer necessary, actively purging its associated representation from memory may be more advantageous to prevent it from interfering with more important, related information that is encountered. In this context, maintenance suppression and other encoding suppression mechanisms (e.g., as engaged by item-method directed forgetting or list-method directed forgetting) may provide a useful means for resolving interference and facilitating new learning.

The field of neuroimaging has seen a gradual increase in sample sizes over time, with the median number of participants rising from 12 before 2012 to 24 by 2018 (76). At the same time, efforts to build large-scale, multisite neuroimaging datasets have further prompted discussions about appropriate sample sizes for neuroimaging research. There is growing consensus that larger-sample studies offer important advantages, including greater statistical power (77), more reliable brain–behavior correlations (78), and improved replicability (79). Nevertheless, smaller sample studies remain valuable, particularly during the early stages of investigating novel paradigms, before scaling up to larger samples. The present study represents an initial step toward addressing the question of how maintenance suppression influences subsequent learning. While we uncovered intriguing patterns in the data, we also acknowledge the limitations posed by our relatively small sample size in this preliminary investigation. To specifically address concerns about statistical power, we computed empirical power estimates (Methods) for each of our major analyses. The underlying assumption is that if reliable effects can be detected with fewer trials and/or participants, this suggests that the study likely had adequate power for those specific effects. Through simulations, we demonstrated that most of our major findings were supported by sufficient empirical power (SI Appendix, Figs. S3, S4, S8, and S9).

In conclusion, our study demonstrates that maintenance suppression, supported by distributed frontoparietal control regions, enhances the encoding fidelity of subsequent information in the hippocampal CA1 region. During retrieval, a neural reactivation bias toward initially learned B items was observed in the maintain condition. In contrast, in the suppress condition, this bias was reversed, favoring the reactivation of more recently learned C items. These findings extend prior evidence on the role of maintenance suppression in resolving memory interference by showing that suppression upregulates the encoding of subsequent information to enhance memory formation and biases retrieval processes to favor newly formed memories following suppression.

Methods

The study was approved by the University of Texas at Austin Institutional Review Board. Informed consent was obtained from all participants. Descriptions of subjects, stimuli, analysis, and statistics can be found in SI Appendix.

Experimental Procedure.

Participants entered the MRI scanner after providing consent and reviewing the instructions. Inside the scanner, they completed six learning-testing cycles (Fig. 1), two anatomical sessions, and six localizer sessions. Throughout all phases, stimuli were displayed on a gray background and projected from the back of the scanner. The experiment was implemented in PsychoPy and lasted approximately 2.5 h.

Learning sessions.

In each of the six learning blocks, participants formed 12 image associations over 12 trials, with each association learned only once. Each trial began with an object image (image A, 2 s, width = 5°, height = 5°) being presented at the center of the screen, followed by a green circle [(0, 0, 255), 4 s, r = 0.5°], indicating that participants needed to memorize the image. In maintenance and suppression trials (4 trials per condition), a second image (image B, 2 s) was then presented. The category of image B (face or scene) was counterbalanced within each condition. In the maintain condition, a green circle appeared after the image, indicating participants to memorize image B (4 s) and create AB associations. In the suppress condition, a red circle [(255, 0, 0)] appeared instead, instructing participants to suppress image B (4 s), with the instruction to “push the image down in their mind every time it comes up, not to empty their minds entirely or to substitute the item in mind with other thoughts”. In baseline trials, a white circle was presented for the entire 6-s period in lieu of a B image. Following this, a third image (image C, 2 s) was presented. The category of image C (face or scene) was counterbalanced within each condition. In the maintenance and suppress conditions, image C was always from the opposite category of image B. The image C was followed by a green circle (4 s), indicating that participants needed to memorize image C and form AC associations. Trials ended with a variable intertrial interval of 2, 4, or 6 s.

Testing sessions.

Following each learning block, participants’ memories of the learned image associations (AB or AC) were tested. Each trial began with a cue image, always an image A from the preceding learning block, being presented at the central top of the screen, 2.5° above the center (8 s). Four candidate probe images, two faces and two scenes, were then shown at the bottom of the screen, 2.5° below the center for 4 s. Participants had to identify which of the four probe images was presented in the same trial as the cue image. To control for general familiarity with probe images, lure images were also chosen from the preceding learning block but not from the same trial as the cue image. Trials ended with a variable intertrial interval of 4 or 6 s. Each learned association was tested only once. That means, in the maintain and suppress conditions, for each object A, either the AB or AC association was tested, but not both. AB and AC associations had an equal probability of being tested. In the baseline condition, only AC associations were tested since no B images were learned. Each learned triplet of objects from the encoding phase was tested once in the testing phase.

The learning-testing cycle was repeated six times, with unique image associations learned and tested in each cycle. In total, 24 AB and 24 AC pairs were learned in the maintenance and suppress conditions, respectively. 24 AC pairs were learned in the baseline condition.

Perceptual localizer task.

Following the main memory tasks, participants completed six runs of perceptual localizer tasks. Each run included 12 trials. In each trial, an image (object, face, or scene) was presented at the center of the screen for 2 s. Participants needed to identify the category of the image and respond with a button press. Each trial ended with a variable intertrial interval of 4, 6, or 8 s. These six functional scans were later used to train a classifier to differentiate neural representations associated with the object, face, and scene categories.

Complementary behavioral experiment.

In each of the three learning blocks, participants formed 24 image associations across 24 trials, with each association presented only once. The trial designs were identical to those in the learning blocks of the fMRI experiment, except that each trial ended with a fixed 2-s intertrial interval. Following each learning block, participants’ memories for the learned image associations (AB or AC) were tested. The test design mirrored that of the main fMRI experiment, with the exception that a test trial ended only after participants made a response by selecting a probe image as their answer.

Supplementary Material

Appendix 01 (PDF)

Data, Materials, and Software Availability

Experimental materials and behavioral data and fMRI data have been deposited in OSF (https://osf.io/7zwd5/) (80) and Zenodo (https://zenodo.org/records/15733963) (81), respectively.