Mood During Consolidation Retroactively Biases Memory for Past Emotional Events

Abstract

Mood-congruent memory (MCM) is a well-established phenomenon that occurs when emotional memory is biased towards content affectively congruent with a past or current mood. To date, however, a majority of research in this area has examined how mood directly impacts the encoding or retrieval of emotional memories, with comparatively less work examining the offline effects of mood during early consolidation. To address this gap in the literature, here we induced happy and sad moods after participants encoded emotional stimuli, and then tested memory at least a day later with both recall and recognition tests. Across multiple experiments (performed 2020–2023) using different stimuli, induction techniques, and memory tasks, we demonstrate that a change in post-encoding mood is associated with mood-congruent shifts in memory for the affective tone of an emotional experience, while also biasing recognition accuracy toward mood-congruent stimuli. These effects were moderated by subjective arousal and self-relevance ratings of the stimuli at encoding, suggesting that phenomenological appraisals of a stimulus shape the influence of mood during consolidation. Our findings propose that MCM can arise from moods dissociated from the initial encoding or subsequent retrieval of emotional material, such that mood retroactively facilitates MCM even after an emotional experience has already occurred.

Emotions motivate and reinforce many aspects of how we behave, both when an emotional experience initially occurs and also when we later remember it. Over time, our memories for emotional experiences can become selective in what is remembered. That is, we may retain only certain details or prioritize specific kinds of memories (e.g., for positive experiences) over others. Moreover, both the frequency and intensity of emotional experiences from one’s recent past can often be overestimated (Miron-Shatz et al., 2009), and current appraisals can shift how one remembers having originally experienced an event. For instance, people’s memories for how they felt during a stressful political election can change depending on how they feel about the candidates now (Levine, 1997; Wilson et al., 2003). Interestingly, how we remember feeling seems to be just as, if not more, predictive of future behavior than the actual experiences themselves (Levine et al., 2009).

Detailing the underlying mechanisms by which emotional memory biases emerge remains an active area of research with direct clinical relevance (Duyser et al., 2020; Duyser et al., 2025). Generally, such biases are made possible by a memory system that is ultimately constructive rather than reproductive. The way by which this memory reconstruction unfolds is influenced by a myriad of factors that shape the content most readily accessible when remembering the past, which often results from what we primarily attended to when initially encoding an episodic event (Aly & Turk-Browne, 2016), but is also influenced by how well that information was stored into memory (Nadel et al., 2012), as well as our goals at retrieval (Kensinger & Ford, 2020). A central adaptive feature of such a constructive system is that it overcomes the need to store an exact representation of every experience into memory. Rather, our memory prioritizes the plethora of information we receive by focusing and retaining details for the most relevant aspects of an event and then reconstructing the missing pieces as needed.

We commonly prioritize and remember the more salient, emotional information in our surrounding environment, often at the expense of mundane or neutral content that is less impactful (Mather & Sutherland, 2011). This emotional enhancement effect on memory has been detailed at both a behavioral and neurobiological level, whereby amygdala-hippocampal interactions enhance consolidation processes under increased arousal (McGaugh, 2004). Emotional memory biases have been shown to take on different forms, sometimes reflecting enhanced accuracy for emotional compared to neutral items (Hamann, 2001), and other times inflated confidence in the recollection of emotional events even without any noticeable differences in memory performance (Phelps & Sharot, 2008). These effects are further moderated by the valence of encoded stimuli, with negative and positive content differentially engaging visuosensory systems in the brain to support sensory recollection at retrieval (Bowen et al., 2018), suggesting that both arousal and valence are key determinants for how emotional information is selectively prioritized in memory. Yet, while these roles of arousal and valence contribute to an understanding of emotional memory formation more generally, the precise mechanisms that subserve biased access to specific emotional events over others remains relatively underspecified. Although decades of research have shown that humans generally remember emotional experiences better than neutral ones, it remains less clear what factors beyond the encoding experience itself can act to further bias the specific emotional experiences that are preferentially retained in memory.

Here we propose that in addition to the initial arousal and valence of an emotional event, our moods during memory consolidation also shape how emotional experiences are later remembered. We hypothesize that moods act as a mnemonic filter for recently encoded material, biasing memory in a mood-congruent direction by strengthening subsequent memory for mood-congruent content compared to mood-incongruent content, and/or shifting appraisals of emotional experiences in a mood-congruent direction. Our approach is informed by a rich literature on mood-congruent memory (MCM), as well as recent empirical demonstrations of retroactive strengthening in memory. We discuss each of these literatures in more detail next.

Mood-Congruent Memory

Affect is often dissociated into two general varieties: emotions and moods. Emotions are characterized as short-lived (transient) reactions to an identifiable stimulus, event, or thought. In contrast, moods are sustained affective states that often develop without awareness of a specific cause or the appraised event has become diffuse and nonspecific, thereby casting a glow or shadow over our thoughts and behaviors (Bower & Forgas, 2000; for further discussion, see also Faul & LaBar, 2023). These two affective experiences are intricately linked, as a series of emotional events can collectively produce a more generalized mood, while a mood can also predispose us to experience emotions congruent with our current affective state (Siemer, 2009). This relationship between emotions and moods extends into memory to produce MCM, the psychological phenomenon whereby memory is biased toward content affectively congruent with a past or current mood (Blaney, 1986).

MCM is thought to emerge from an associative network, whereby activating a mood spreads activation to closely linked, emotionally congruent content (Blaney, 1986; Bower, 1981). When moods are present during encoding, this spreading activation facilitates emotional responding to mood-congruent stimuli and, consequently, provides a richer encoding experience for those items. Alternatively, when moods are present at retrieval, they increase the availability of mood-congruent content during the memory search process, especially when retrieval cues are sparse. Decades of empirical research have supported these general claims, although the exact neurocognitive mechanisms by which MCM occurs remains unclear. A recent systematic review from Faul & LaBar (2023) reported consistent evidence for MCM across laboratory studies using mood inductions over the past four decades, supporting the notion that moods are a significant factor governing how specific emotional memories are prioritized over others. This review highlighted work showing that moods not only influence what is retained in memory (e.g., via recognition tests), but also how an event is reconstructed during recall, such as shifting the valence of recalled memories in a mood-congruent direction (Drače & Desrichard, 2013). Studies have demonstrated that mood-related effects are most prominent among self-referential content (e.g., Itoh, 2004; Nasby, 1996) – confirming the conclusion of earlier reviews (Blaney, 1986) – suggesting that the personal relevance of an event influences its positioning within an associative network of affective information, as theorized by Bower (1981). Additionally, other studies have shown that the specificity of mood-emotion congruency (Bland et al., 2016; Hansen & Shantz, 1995) and the intensity of the induction response (Drače, 2013; Drače et al., 2015) further strengthen MCM.

However, the systematic review by Faul & LaBar (2023) also noted significant methodological gaps in our understanding of MCM that need to be addressed, such as the delay between encoding and retrieval. Because most MCM studies test memory almost immediately after encoding, they cannot assess the influence of longer-term consolidation processes that are known to facilitate emotional memory biases (Sharot & Phelps, 2004; Sharot & Yonelinas, 2008). According to the mediation theory of emotional memory, immediate enhancements in memory are primarily governed by organization, distinctiveness, and attentional factors at encoding that facilitate retention for arousing information (Talmi, 2013). After a delay, memory for emotional content is then shaped by separate consolidation processes involving amygdala-hippocampal interactions, which may help to explain relatively weak correlations between immediate and delayed emotional memory effects (Schümann et al., 2018). When considering this distinction in immediate versus delayed memory mechanisms alongside the limited research on delayed MCM, it’s possible that mood primarily governs the emotional content immediately retained in memory (by exerting an influence on attentional processes), with comparatively less influence on consolidation mechanisms. In other words, a majority of the MCM literature does not address any long-term effects of mood on memory, and thus cannot rule out the possibility that mood only indirectly influences memory by shaping attentional mechanisms at encoding and/or retrieval (Faul & LaBar, 2023). Alternatively, here we propose that one’s mood can directly affect the consolidation of emotional material into long-term memory, although testing this proposal not only requires measuring MCM after a sufficient delay for consolidation to take place, but also isolating the effects of mood on early stages of memory consolidation.

Relatedly, then, a prominent theoretical gap in the literature is whether MCM can emerge from moods detached from the initial encoding or subsequent retrieval of emotional information. Theoretically, the basic tenets of associative network theory suggest that post-encoding mood can spread activation to recently linked content in memory, thereby retroactively strengthening mood-congruent items during consolidation. However, nearly all studies on MCM have induced mood immediately prior to the encoding and retrieval tasks (Faul & LaBar, 2023). These inductions shift participants’ attentional focus to mood-congruent information and thus, as mentioned, the resulting MCM effects may be attributable to such attentional biases. Moreover, studies that induce mood immediately prior to encoding or retrieval produce other mood-related effects that can confound influences on memory. Positive moods, for instance, broaden the scope of visuospatial attention and facilitate top-down assimilative processing, whereas negative moods promote bottom-up accommodative processing that narrows attention (Fiedler, 2001), both of which may fundamentally change how participants engage with an encoding or retrieval task. Even though a few studies have attempted to apply post-encoding manipulations to shift affect (Liu et al., 2008; Nielson & Lorber, 2009; Wang & Ren, 2017), these studies evaluated nonspecific increases in aversive arousal (an elevated level of physiological stimulation directed toward an unpleasant stimulus), rather than the effects of targeted positive or negative mood inductions, per the definition of mood provided earlier (see Faul & LaBar, 2023). It therefore remains unclear whether MCM can emerge from mood induced after encoding, though testing such an effect is a vital next step in improving our understanding of the mechanisms subserving MCM.

In summary, support for MCM is constrained primarily to same-day memory effects and moods that immediately precede the encoding or retrieval of emotional material. These limitations impact the generalizability of such findings to real-world scenarios, where moods may have longer-term consequences for how an emotional past is remembered, even if they do not co-occur with emotional experiences. This proposal reflects some aspects of tag-and-capture models of memory, which suggest that strong experiences can retroactively bolster memory for weaker experiences by increasing activity in the same neural ensemble and stabilizing the weaker memory trace (Dunsmoor et al., 2022; Ritchey et al., 2016). Unlike typical studies on this phenomenon, however, here the memoranda are not weak/neutral but rather are emotional in nature. While these items tend to be inherently prioritized in memory, there is usually competition at encoding among different emotional categories or dimensions (e.g., happy vs. sad). A post-encoding mood manipulation may generate an affective state that prioritizes one of these emotional constructs, which serves to filter emotional material congruent with that state for consolidation into long-term memory. Although mood induction studies have yet to reveal retroactive influences of induced moods, a growing body of work with post-encoding conditioning tasks have demonstrated such effects, which we turn to next.

Selective and Retroactive Strengthening of Memory for Related Events

Outside of the MCM literature, several studies have shown that emotional events taking place after encoding can retroactively strengthen the consolidation of conceptually related information. For example, Dunsmoor et al. (2015) asked participants to encode a series of neutral images depicting either animals or tools, followed by a fear-conditioning task with new images of animals and tools where one category was reinforced with shocks while the other category was explicitly unreinforced. Memory for items before conditioning that were conceptually congruent with the subsequently reinforced category exhibited enhanced memory compared to the unreinforced category of images, but only if memory was tested 6 hr later or the next day and not immediately after (Dunsmoor et al., 2015). Thus, initially neutral image sets shifted in their priority depending on new information gained after encoding, although such effects necessitated a sufficient delay for consolidation processes to occur. Similar findings have been shown not only in other post-encoding fear-conditioning manipulations (Hennings et al., 2021), but also in post-encoding reward manipulations (Braun et al., 2018; Patil et al., 2017) suggesting that both negative and positive post-encoding experiences can capture recently encoded materials and enhance their consolidation into memory (but see Kalbe & Schwabe, 2022).

Additionally, exposure to novelty after an initial learning experience can improve long-term memory for what was previously learned (Ballarini et al., 2013; Ramirez Butavand et al., 2020). While such effects have only recently been tested in humans, rodent research has shown that novelty exposure promotes retroactive enhancements by selectively promoting plasticity-related protein synthesis in the hippocampus (Moncada & Viola, 2007; Salvetti et al., 2014). Moreover, evidence suggests that post-encoding stress also selectively acts upon recently learned information, such that stress prioritizes memory for events eliciting higher amygdala and hippocampal activity during encoding (Ritchey et al., 2017) or information learned in the same environment (e.g., the same room; Sazma et al., 2019), and in some cases enhances the mnemonic discrimination of negative content specifically (Cunningham et al., 2018, 2021). Thus, whether in the context of stress, novelty, reward, or fear conditioning, it is theorized that post-encoding experiences can reactivate shared neural ensembles to selectively bolster consolidation processes for previously-experienced content (Dunsmoor et al., 2022).

We propose that one’s mood after encoding may also selectively bias memory for recently experienced events, although such effects would emerge from affective congruency rather than conceptual congruency. Examining retroactive strengthening in the context of mood, however, poses a number of new challenges compared to previous investigations. Namely, inducing a mood is fundamentally distinct from conditioning tasks that associate stimuli with threat or reward, or from novelty manipulations where a completely new learning experience takes place. Instead, testing the effects of mood requires inducing an affective state that is both specific and strong enough to overlap with the emotionality of previously encoded stimuli, although even then the degree of congruency will vary depending on the efficacy of the manipulation across individuals. Moreover, in contrast to other post-encoding manipulations that have commonly tested memory for neutral stimuli, examining MCM necessitates having participants encode emotional stimuli, and thus memory is influenced by the original emotional experiences themselves and must be appropriately accounted for in analyses.

The research literature provides conflicting support for the possibility of MCM resulting from post-encoding mood. As mentioned, the associative network theory for MCM theoretically suggests that mood, irrespective of timing, can spread activation to linked affective material, thereby strengthening memory even for previously seen mood-congruent content. Yet, for several reasons, it remains unclear whether this is actually the case. First, despite over four decades of research on MCM, such an effect has yet to be empirically shown (for further review, see Faul & LaBar, 2023). Second, other models of mood and memory, such as the dual-force model (Fiedler, 2001), propose that moods promote characteristically different processing styles, with negative moods supporting accommodative processing that facilitate the conservation of stimulus representations (i.e., enhanced attention to stimulus-specific details). It is possible that negative moods after encoding could generally improve memory compared to positive moods, rather than selectively targeting only negative content (or being sensitive to other properties of a recently encoded stimulus, such as arousal or self-relevance). Alternatively, as we propose here, these different processing styles may have less influence on memory if mood is experienced during offline consolidation, given that previously encoded material is no longer the focus of one’s attention.

Third, despite the aforementioned fear, reward, and novelty studies illustrating retroactive effects on memory, such findings have not been consistently observed. A recent meta-analysis concluded no reliable evidence for selective retroactive enhancements when primarily assessing post-encoding fear and reward manipulations, although the relatively small number of experiments may be a limiting factor, as well as the potential influence of moderating variables that are understudied (Koevoet & Postma, 2023). Fourth, a core proposal from tag-and-capture models is that the initial material is weakly encoded, thereby allowing for a salient post-encoding event to release plasticity-related proteins that the weakly stimulated synaptic pathway can now capture (Dunsmoor et al., 2022). This threshold for weak encoding remains ambiguous, as well as the relative strength of the post-encoding experience, but emotionally arousing content may nevertheless have too strong of an encoding signature to further benefit from any post-encoding events. Alternatively, as we propose here, it may be possible that emotionally arousing material can still be selectively filtered in memory, but that such an effect necessitates a more durable change in affect (i.e., mood) after encoding. Indeed, in line with associative network theory (Bower, 1981), emotional experiences with higher arousal and/or self-relevance may be more closely integrated with the associated nodes of a congruent mood state, and thus more readily affected by an induced mood during consolidation.

The Present Research

Here we designed a novel set of experiments to test our hypothesis that mood induced during early consolidation (shortly after encoding) retroactively biases memory for recently encoded emotional content. Across multiple experiments using different encoding tasks, retrieval tasks, stimuli, and induction methods, we tested if post-encoding mood influences not only what people remember from the past, but also how they reconstruct their affective experiences. To this end, in Experiments 1A and 1B we asked participants to imagine hypothetical scenarios with emotional outcomes during encoding, induced a happy or sad mood shortly thereafter using different stimuli, and then tested their memories for these imagined events at a subsequent retrieval session. In Experiment 2, we examined whether post-encoding mood can also retroactively bias recognition memory for emotional pictures.

We hypothesized that post-encoding mood biases next-day memory in a mood-congruent manner by shifting emotional scenarios to be misremembered as sadder or happier than they originally were, when having been in a sad or happy mood, respectively, during early consolidation. We further hypothesized that recognition accuracy would be enhanced for mood-congruent versus mood-incongruent stimuli, such that emotional stimuli (both imagined scenarios and visual images) are best remembered when having been in a congruent mood after encoding. Importantly, we expected these effects to be moderated by phenomenological attributes that influence how stimuli are organized within an associative network of emotional information. These attributes include valence, which characterizes the affective congruency of a stimulus with the post-encoding mood, as well as arousal and self-relevance. We suspected that more arousing and/or self-relevant stimuli would already be most prioritized and integrated into memory networks during early consolidation and therefore most affected by the spreading activation of mood during this time. Yet, we also acknowledged the possibility that such moderation may only emerge for a specific valence, or these attributes could influence overall memorability without necessarily shaping the sensitivity of stimuli to post-encoding mood change. To test these possibilities, we evaluated models that allowed for moderation effects and, if supported by the data, unpacked lower-level interactions and simple slopes to determine where MCM effects were most evident.

Experiments 1A and 1B

For our first set of experiments, we designed the study tasks and materials based on observed moderators of MCM and methodological recommendations provided by Faul & LaBar (2023). First, regarding the choice of stimulus and encoding task, MCM research has consistently shown that self-referential encoding enhances the effects of mood on memory. Accordingly, for Experiments 1A and 1B we used hypothetical text-based scenarios as stimuli to allow participants to construct imagined, personal events. For each scenario, we asked participants to provide phenomenological ratings of their imaginative experience with respect to its valence, arousal and self-relevance. Using imaginative scenarios as stimuli also allowed participants to engage in a generative and elaborative task, which has been shown to be more susceptible to mood-related effects (Blaney, 1986; Bower, 1981; Forgas & Eich, 2012).

Regarding the mood induction procedures, we designed inductions that exposed participants to combinations of stimuli with similar affect. In Experiment 1A, participants recalled a series of autobiographical memories while listening to mood-congruent music. In Experiment 1B, participants recalled one autobiographical memory while also listening to mood-congruent music, and then watched a collection of real-life news and film clips. We used these multimodal techniques to induce moods that were less specific to just one individual stimulus, as this approach appreciates the distinction between moods and emotions (e.g., moods being more diffuse and non-specific) and overcomes potential limitations of previous studies that induced affect with only a single stimulus (which may have elicited transient emotional responses rather than a durable shift in affective state).

Regarding the retrieval task, research suggests that moods have a stronger influence on memory recall than recognition, given that sparse retrieval cues place greater reliance on participants to reconstruct a memory trace from internal contextual resources such as mood (Singer & Salovey, 1988). However, numerous studies have also documented MCM with recognition tests (Faul & LaBar, 2023). We therefore designed a retrieval task that initially presented participants with only partial cues of the scenarios they had imagined during encoding (by removing the emotional outcome of the event from the text) and then asked participants to recall how they remember originally experiencing each scenario (i.e., remembered valence, arousal, and self-relevance) based on their recall of the full event. The recall task was then followed by a recognition task where participants were given the opportunity to complete the full scenario from a list of provided options.

We first collected data on Experiment 1A as an initial test of the feasibility of our study tasks and the efficacy of a post-encoding induction, as well as part of a separate examination of psychophysiological responding to imagined events and autobiographical memory recall that is outside the scope of the present paper (Faul et al., 2023). To increase our power to test the proposed interactive effects on memory – as well as the generalizability of these effects across different contexts, stimuli, and induction techniques – we also conducted Experiment 1B with a sample of online participants who completed the same encoding and retrieval tasks as Experiment 1A but with a different induction method. This approach more than doubled the final sample size used in all memory analyses. We performed memory analyses on the pooled data using linear mixed effects modeling of all available observations (total trials = 7,739) while accounting for variability in the data across participants, stimuli, and experiments.

Materials and Methods

Participants

In Experiment 1A (January 2020 – January 2022), 98 participants volunteered for the study and completed written informed consent in accordance with the Duke University Institutional Review Board guidelines. Inclusion criteria included an age range of 18–39 years old, as well as no known history of a neurological condition or psychiatric disorder. Fourteen participants received course credit for participating, while the remaining participants received monetary compensation ($15/h). Twelve participants withdrew from the study early due to scheduling issues. Moreover, six participants were excluded from analyses due to outlier behavior (> 3 SD) reflecting low accuracy on the active fixation task at the encoding or retrieval sessions (n = 3; this exclusion was applied to remove inattentive participants with unreliable data), or high levels of negative mood prior to the mood induction and/or prior to the retrieval task (n = 3; this exclusion was applied to minimize any mood-related effects not associated with the induction). Applying these exclusions resulted in a sample of 80 participants (48 women, 32 men; M_age = 21.3 yrs, SD = 3.4; 17 Hispanic or Latino, 62 not Hispanic or Latino, 1 unknown; 1 American Indian or Alaska Native, 28 Asian, 8 Black, 35 White, and 8 other or unknown), equally split between the happy and sad groups via random assignment. For the pooled memory analyses, we further excluded two participants with word recognition accuracy below 50%, indicating unsuccessful encoding that impedes the estimation of mood-related influences due to atypical floor effects (although note that our findings remain consistent when these participants are still included in analyses; see Supplemental Material for these sensitivity analyses). This approach also allowed us to combine more homogenous samples across Experiments 1A and 1B.

In Experiment 1B (December 2021), all participants were recruited via the Prolific crowdsourcing platform and received monetary compensation for their time ($10 per session). Inclusion criteria included an age range of 18–39 years old, as well as no known history of a neurological condition or psychiatric disorder. All participants completed written informed consent in accordance with the Duke University Institutional Review Board guidelines. 165 participants were paid for completing the encoding session, while 123 participants were paid for returning and completing the retrieval session. We only analyzed data from participants who completed both sessions. Of the 123 participants who completed retrieval, 9 participants were excluded for indicating that they were honestly not paying attention during the encoding or retrieval tasks (our final attention check question; see Supplemental Materials for more information). We further excluded 7 participants due to outlier behavior (> 3 SD) reflecting low accuracy on the active fixation task at the encoding or retrieval sessions (n = 5) or high levels of negative mood prior to the mood induction (n = 2). We also excluded one additional participant who spent nearly an hour on the video portion of the induction task (indicating inattention or severe buffering issues with the videos). Applying these exclusions resulted in a sample of 106 participants (53 women, 51 men, 2 non-binary; M_age = 28 yrs, SD = 6.1; 7 Hispanic or Latino, 99 not Hispanic or Latino; 1 American Indian or Alaska Native, 7 Asian, 10 Black, 80 White, 7 more than one race, and 1 unknown), with 52 participants randomly assigned to the happy mood group and 54 participants to the sad mood group. For the pooled memory analyses, we further excluded 12 participants with recognition performance below 50% accuracy.

The novelty of our experimental method – as exemplified by a post-encoding induction but also choice in stimulus materials, memory tests, and mixed-effects analyses that are distinct from existing literature (Faul & LaBar, 2023) – precluded estimation of a precise effect size for sample size estimation. Therefore, we based our sample size for Experiment 1A on previous demonstrations of post-encoding biases from Dunsmoor et al. (2015) who demonstrated retroactive memory enhancements in a sample of 30 participants and reported a Cohen’s d_z effect size of 0.45, which suggests at least 41 participants are required to detect a similar effect with 80% power using a two-tailed paired t-test (Kalbe & Schwabe, 2022). Moreover, using G*Power 3.1 (Faul et al., 2009), we separately calculated that an estimated sample size of 54 participants (27 per group) is needed to achieve 80% power to observe a within-between interaction (e.g., Time x Group effect) for the mood induction response, assuming a large effect size of η_p² = 0.14 based on past literature (Ferrer et al., 2015). This is only an approximation, however, given that standard power analyses cannot incorporate many study-specific influences that may shape the generalizability of the observed effects (e.g., the choice of stimuli and the presentation speed of those stimuli). We therefore planned to collect eighty participants for Experiment 1A, with forty participants per happy and sad induction group. As mentioned, we then increased the size of our total sample by also collecting data from the same encoding and retrieval tasks in Experiment 1B. Combining the samples from Experiment 1A and 1B provided sufficient power to detect a within-between interaction (e.g., Condition x Group effect on memory) even when assuming a small/medium effect size of η_p² = 0.04, as such an effect requires 120 participants to achieve 80% power (using the as in SPSS effect size specification). Note, however, that in all memory analyses we used post-encoding mood change as a continuous measure (instead of group assignment) as well as individual ratings of valence, arousal, self-relevance, and confidence that increased the sensitivity of our analyses compared to just comparisons across categories. By pooling data from both experiments, we were able to build models from all available observations across a total of 172 participants after exclusions for low memory performance (96 women, 74 men, 2 non-binary; M_age = 24.8 yrs, SD = 6.0; 84 participants in the happy mood group and 88 participants in the sad mood group).

General Procedure

A general overview of the experimental design is provided in Figure 1. Both experiments consisted of an encoding task and post-encoding mood induction on Day 1, and a next-day retrieval task on Day 2. Experiment 1A was performed in our laboratory space, while Experiment 1B was performed online. The two experiments primarily differed in the mood induction technique that was used, but otherwise were nearly identical in design.

Figure 1. Experimental design for Experiments 1A and 1B.

Note. Mood ratings were acquired before encoding, after encoding/before the induction, after the induction, and before retrieval. Encoding consisted of imagining and rating emotional scenarios based on valence, arousal, and self-relevance (in a random order). Retrieval consisted of viewing those same scenarios but with critical words omitted from the text, and participants were asked to recall the ratings they had provided the day prior, as well as pick the correct word that filled in the blank.

Encoding consisted of reading and imagining hypothetical scenarios, as well as providing ratings of valence, arousal, and self-relevance for each scenario (see the Encoding Task section below for further details). Then, after a brief delay, participants completed the mood induction (see the Mood Induction section for more details). The next day, participants returned to complete the retrieval session. In Experiment 1A, the retrieval session was scheduled for the same time of day as the encoding session. In Experiment 1B, the retrieval session was posted online in the morning and participants started the retrieval session, on average, 29.6 hours (SD = 3.8) after the encoding session. The retrieval task consisted of a surprise memory test that involved viewing the scenarios from the day prior with critical words omitted from the text (more details on this task are provided below in the Retrieval Task section). Additional procedural details are provided in Supplemental Materials.

Mood Ratings

For Experiment 1A, mood was measured with the abbreviated Profile of Mood States (Grove & Prapavessis, 1992), which asks participants to rate how they are feeling in the moment on a scale of 0 (not at all) to 4 (extremely) for various affect-related adjectives. Ratings from these items were combined to produce seven mood subscales that measure tension, anger, fatigue, depression, esteem-related affect, vigor, and confusion. The original POMS included 40 items, but to obtain a more comprehensive measure of positive mood we added an additional five positive adjectives to the questionnaire to create a general positive affect scale: “content”, “hopeful”, “happy”, “clear-headed”, and “relaxed”. This modification has also been made in prior work (Faul & De Brigard, 2022; Pinheiro et al., 2013). A total mood score was calculated by subtracting the sum of all negative subscales from the sum of all positive subscales.

For Experiment 1B, mood was measured with the Brief Mood Introspection Scale (Mayer & Gaschke, 1988), which asks participants to rate their present mood on a scale of 1 (definitely do not feel) to 7 (definitely feel) for 16 different affect-related adjectives. We also asked participants to provide an overall mood rating from −10 (very unpleasant) to 10 (very pleasant). We decided to use the shorter BMIS instead of the POMS to prevent against survey exhaustion, given that Experiment 1B participants completed a series of trait questionnaires at the beginning of Day 1 (whereas Experiment 1A participants did so at a separate session a week prior), and we were not present to ensure appropriate completeness of all survey items. We constructed a total mood metric by combining the overall mood rating with specific ratings for happiness and sadness ([Overall + Happy] – Sad), which closely mirrored the distribution of total mood ratings from Experiment 1A. Higher values reflect more positive, happy moods, whereas lower values reflect more negative, sad moods.

Stimuli

Encoding stimuli consisted of hypothetical scenarios written in the second-person perspective, originally created by Fields & Kuperberg (2012). All scenarios consisted of two sentences, with the second sentence containing a critical word that determined whether the outcome of the event was happy, neutral, or sad. For instance, “At your job, all the employees get yearly reviews. Your evaluation is quite encouraging/succinct/discouraging this year.” Previous use of these scenarios has shown that they successfully elicit self-referential and discrete emotional responses (Faul et al., 2022; Fields & Kuperberg, 2012, 2014, 2016). Prior to conducting the present experiments, a stimulus set was chosen based on a comprehensive norming analysis performed on data acquired from separate online raters (see Supplemental Materials for more detail). Based on these normative ratings, 45 imaginative scenarios were selected for each experiment, with 15 scenarios assigned to each normative emotion category (sad, neutral, and happy). We only chose one variant for each scenario (e.g., the encouraging outcome in the above example) in order to utilize stimuli that elicited the strongest normative ratings of happiness and sadness.

In both experiments, scenarios were selected such that all three normative categories differed in valence (ps < .001), while sad and happy scenarios were equated on arousal but both more arousing than neutral scenarios (ps < .001). Among the online raters in our norming assessment, sad scenarios were also deemed as less self-relevant than happy and neutral scenarios, reflecting a general positivity bias in perceived self-relevance. For each experiment, all participants imagined the same set of scenarios during encoding, but different sets were used in Experiment 1A and 1B (see Supplemental Materials).

Due to the variant format created by Fields and Kuperberg (2012), each scenario that we showed during encoding was associated with two possible variants based on replacement of the critical word that shifted the emotionality of the outcome (e.g., succinct and discouraging in the previous example). The critical words from these variants were used during the recognition task, which required participants to select the exact word that completed each scenario they had imagined at encoding (see below for further details in the Retrieval Task section). To increase the difficulty of the task, we added three additional lure words that were each synonymous with the critical words from the original three variants, resulting in a total of six possible word selections during the recognition task: the original critical word of the scenario shown at encoding, a same-valenced lure, and two words from each of the other two variants. For example, a sad scenario used in the encoding stimulus set was affiliated with six total critical words shown during the recognition task: the sad word that originally completed the scenario at encoding, a sad lure word, two neutral words, and two happy words. Importantly, this approach increased the difficulty of the recognition task by requiring participants to not only remember the general emotional outcome of the scenario they had imagined, but also the specific word that determined the outcome. We ensured that these word selections did not significantly differ across the emotion categories in length, concreteness, and frequency of use in the English language (see Supplemental Materials for more information on the selection of additional critical words).

Encoding Task

The encoding task consisted of reading and imagining all 45 scenarios from the normed stimulus set. Scenarios were presented in a pseudorandomized order, with every three trials consisting of one happy, one sad, and one neutral scenario randomly arranged. As such, scenarios of the same normative category were never repeated more than once. On each trial, participants were asked to read and imagine the depicted event, which was shown in the center of the screen for 16 seconds in Experiment 1A and 12 seconds in Experiment 1B (we shortened the length of this period given that we were not collecting psychophysiological data and only behavioral responses in Experiment 1B). Halfway into this imagination period the words “I feel…” appeared on the screen to remind participants to focus on how they are feeling as they imagine the scenario and to prepare for the upcoming ratings. The imagination period was followed by ratings of valence (1-very sad to 7-very happy), arousal (1-very calm to 7-very intense) and self-relevance/likelihood of the depicted event occurring in real life (1-very unlikely to 7-very likely). These three ratings were arranged in a random order and each displayed for four seconds. All ratings were provided by participants with keys 1–7 on the keyboard. Trials were separated by an active fixation task that required participants to identify a number as even (G key) or odd (H key) within a 2.5-second window, and an additional jittered interval of 4.5 seconds (Experiment 1A) or one second (Experiment 1B) with a fixation cross on the screen before the presentation of the next scenario.

Mood Induction

In Experiment 1A, happy and sad moods were induced using autobiographical memories that were provided by participants at an online session 6–8 days prior. Participants had provided 15 happy and 15 sad memory cues at this previous session, but were told prior to the mood induction that only the happy or sad memories would be displayed (although they were not told which one of these would occur). Memories were randomly sorted and presented in succession, mirroring the presentation format of the encoding task. For each trial, a memory cue was presented on the screen for a total of 20 s, with the additional cue “I feel…” appearing 10 s into the presentation of the memory cue. Thereafter, participants provided ratings of valence and arousal (4 s each) in the same manner as they had for the scenarios. Trials were separated by a passive fixation cross (5 – 6 s). Mood-congruent music played throughout the entirety of the induction period, consisting of three happy or three sad instrumental songs that were arranged in a random order (see Supplemental Materials for norming of music clips). The induction task lasted for approximately 10 minutes.

Experiment 1B used a different induction task than Experiment 1A, given that happy and sad autobiographical memories were not collected a week prior to the encoding session. The induction began with instructing participants to recall one sad memory or one happy memory (depending on group assignment), for which they provided a few sentences explaining the event and providing the approximate date. Thereafter, participants completed questions drawn from the autobiographical memory questionnaire (Fitzgerald & Broadbridge, 2013) and the memory experiences questionnaire (Sutin & Robins, 2007), which included rating the memory’s characteristics in terms of clarity, color, visual detail, sound, smell, touch, taste, vividness, composition, location, spatial arrangement of people and objects, time, valence and arousal of the emotions felt then and now, frequency of talking and thinking about the event, and visual perspective. These ratings were collected primarily to ensure that participants were elaborating on the sad or happy event, which on average lasted for 4.5 minutes (SD = 2.5). During this time, the same mood-congruent music used in Experiment 1A played while participants completed the ratings. Once the memory component of the induction was complete, participants next watched a series of happy or sad news clips (depending on group assignment). The news clips were originally acquired from Samide et al. (2020), depicting real-life news stories of positive and negative events. We normed a subset of these videos to select five separate news clips for each group that were specific for evoking feelings of sadness and happiness, given that the original study did not collect ratings for specific emotion categories (see Supplemental Materials for an overview of our norming approach). The final component of the induction involved viewing a sad or happy film clip. The sad film clip was drawn from the movie The Champ, depicting a child reacting to the death of his father, which has been shown to be specific in evoking feelings of sadness (Gross & Levenson, 1995). The happy film clip was drawn from the movie Wall-E, depicting two robots dancing in space, which has been shown to be specific in evoking feelings of happiness (Gabert-Quillen et al., 2015). To ensure the specificity and strength of these film clips for the sad and happy inductions, we also obtained normative ratings from the same sample of online participants who rated the news clips (see Supplemental Materials).

Retrieval Task

During retrieval, participants were presented with all 45 scenarios they had seen the day prior. Importantly, the critical word that shaped the emotionality for each scenario was omitted from the text, thereby requiring participants to recall the emotional outcome of the event they had imagined the day prior (e.g., “At your job, all the employees get yearly reviews. Your evaluation is quite _________ this year.”). On each trial, participants were provided 8 seconds in Experiment 1A and 6 seconds in Experiment 1B to read the scenario, with a blank space representing the location of the critical word (participants were told that this spacing was the same across all scenarios and thus not indicative of the length of the actual word). Thereafter, participants were asked to indicate the corresponding valence, arousal, and self-relevance ratings they had provided the previous day, based on however they remember the full scenario. The partial scenario text remained at the top of the screen during this time. For each of the recall ratings (presented in a random order) participants were provided 7–8 s to make the recall rating as well as the corresponding confidence in their response (1-not at all confident to 7-very confident). After completing all the recall ratings, participants were next shown six possible critical words that could complete the scenario. As mentioned in the Stimuli section, these six words consisted of the actual critical word that was shown at encoding, a same-valenced lure, and four words from the other two normative emotion categories. The six words were presented in a circle, with each word placed equidistant from the center of the screen. Participants were provided eight s to choose a word, and an additional 3–4 s to rate their confidence in the selection. Each trial was separated by the same even/odd active fixation task that was used at encoding. To prevent against implicit memory responses associated with the response key mapping at encoding, we coded the retrieval task in both Experiment 1A and 1B so that the participants could simply click on the screen to make their responses for all components of the task instead of a key press. Scenarios were again presented in a pseudorandomized order, such that every three trials consisted of one happy, one sad, and one neutral scenario arranged in a random sequence. The task was split into three blocks, with each block consisting of 15 scenarios.

Statistical Analyses

All statistical analyses for both Experiments 1A and 1B were performed in R. ANOVAs (Type III sum squares) of mood induction response were computed with the afex package (version 1.3.0; Singmann et al., 2023), and follow-up comparisons with the emmeans package (version 1.8.7; Lenth, 2016; Lenth et al., 2023). We used Greenhouse-Geisser sphericity correction for degrees of freedom, and follow-up comparisons are reported after Bonferroni correction for multiple comparisons. Encoding and retrieval analyses of phenomenological ratings and recognition accuracy used linear mixed-effects models (LMMs), specified with random intercepts for each subject and each stimulus. In the pooled analyses, even though some scenarios were used across both experiments, we still coded them as distinct items in the data given that they were experienced in different contexts and within overall different stimulus sets (i.e., scenarios were nested within experiment). For the pooled memory analyses, we also included a random slope of post-encoding mood change for each scenario to account for potential differences in the sensitivity of stimuli to mood-related biases. LMMs were tested using the lme4 package (version 1.1.34; (Bates et al., 2015) and lmerTest package (version 3.1.3; (Kuznetsova et al., 2017), and significance for fixed effects were assessed using Satterthwaite approximations to degrees of freedom. To unpack the observed higher-level interactions, the emmeans package was used to compute joint tests, estimated marginal means, and simple slopes at lower levels. That is, all effects were examined by extracting estimated marginal means from the full model instead of running new models on subsets of the data (Garofalo et al., 2022). For all of the reported LMMs, the variance inflation factor for each main effect term (still accounting for all higher-level interactions and quadratic terms) ranged from 1–6 (computed using the car package, version 3.1.2; Fox & Weisberg, 2019), which is considered an acceptable range and suggests the absence of multicollinearity (O’Brien, 2007). All of the reported LMMs converged and did not have singular fits.

Examining the influence of post-encoding mood on next-day memory retrieval requires accounting for encoding-related variables that may influence susceptibility to MCM (i.e., phenomenological attributes such as valence, arousal, and self-relevance ratings), as well as memory strength at retrieval (i.e., confidence ratings). Our analytical approach was therefore designed to consider these variables jointly, allowing us to examine the combination of factors that best predict MCM. For our primary two models of interest at recall, we examined remembered valence predicted by the interaction of confidence, post-encoding mood change (modeled as a continuous measure) and each of the other two phenomenological attributes from encoding (arousal and self-relevance). We z-scored the encoding ratings across participants to preserve scale interpretability (and because stimuli were specifically normed on these dimensions), but z-scored the confidence ratings within-participant to account for any baseline differences at retrieval. Note, however, that similar effects were observed regardless of centering choice. Valence ratings were modeled via a quadratic function to capture nonlinear relations reflective of asymmetric MCM.

Recall Analysis.

To identify the optimal model structure while retaining model interpretability and reducing the risk of overfitting, we compared two possible higher-order models that allowed for interactions among mood, confidence, and encoding features, differing only in which phenomenological attribute (arousal or self-relevance) served as an additional moderator. The model with arousal as a moderator yielded a substantially better fit based on Akaike Information Criterion (AIC; delta = 30.81), and log-likelihood (delta = 15.41) comparisons, and therefore was selected as our primary model of interest. This model also yielded a significantly better fit than any simpler models with fewer terms or interactions (all p < .001 via likelihood ratio tests), confirming the importance of higher-level interactions among our hypothesized variables of interest. Our final model consisted of the following:

$$\text{Model}1:\text{Remembered-Valence}~\text{Mood}*\text{Confidence}*\text{Encoding-Valence}*\text{Encoding-Arousal}+\text{Experiment}+\left(1|\text{Subject}\right)+\left(1+\text{Mood}|\text{Scenario}\right)$$

In follow-up analyses, we allowed Experiment to interact with all other variables to confirm the stability of effects across experiments. We note that in exploratory assessments, we similarly evaluated the remembered arousal and self-relevance of scenarios as dependent variables. However, these models did not reveal stable effects across experiments and thus were not further examined.

Recognition Analysis.

For the recognition analysis we tested generalized LMMs, given that accuracy was coded as a binary outcome. We still included a random intercept for each subject and scenario, but a random slope of post-encoding mood change for each scenario was dropped due to singular fit. Using the same approach as above, we now found that the model with self-relevance as a moderator produced a better fit than arousal (AIC delta = 8.45; log-likelihood delta = 4.23). We also confirmed that this model yielded a better fit when compared to simpler models with fewer terms or interactions (all p < .03 via likelihood ratio tests), except for a model without the quadratic effect of valence (p = 0.365). Thus, in final analyses, we only included the linear effect of valence. Our final model therefore consisted of the following:

$$\text{Model}2:\text{Recognition Accuracy}~\text{Mood}*\text{Confidence}*\text{Encoding-Valence}*\text{Encoding-Relevance}+\text{Experiment}+\left(1|\text{Subject}\right)+\left(1|\text{Scenario}\right)$$

Transparency and Openness

We report how we determined our sample size, all data exclusions, all manipulations, and all measures in the study. The de-identified data and analysis code on which the results and conclusions are based are available at the following Open Science Framework website (Faul & LaBar, 2025): https://osf.io/m3nz7/. This design and its analysis were not pre-registered.

Results

Encoding

An overview of ratings acquired during encoding is provided in Figure S6. We observed substantial variation in the subjective emotional experience reported by participants, which we leveraged in subsequent memory analysis using LMMs to examine which combinations of phenomenological attributes were most sensitive to MCM. Nevertheless, although our memory analyses used individual valence ratings for each stimulus instead of category assignment, we still examined whether the ratings from our participants during encoding matched with the category assignments from our normative data to ensure that stimuli were appropriately selected and generally comparable across experiments.

Indeed, the encoding ratings replicated the normative ratings and were also very similar across both experiments, with significant main effects of emotion category observed for ratings of valence (Exp. 1A: F_2,44.951 = 255.717, p < .001; Exp. 1B: F_2,44.986 = 343.449, p < .001), arousal (Exp. 1A: F_2,44.724 = 23.610, p < .001; Exp. 1B: F_2,44.892 = 46.382, p < .001), and self-relevance (Exp. 1A: F_2,44.905 = 6.754, p = .003; Exp. 1B: F_2,44.836 = 14.337, p < .001). Follow-up pairwise comparisons (see Table S1 for a report of all statistics) indicated that, in both experiments, happy, neutral, and sad scenarios all differed in valence ratings. Happy and sad scenarios were also rated as more arousing than neutral scenarios, with no difference between happy and sad scenario arousal ratings in Experiment 1A, although sad scenarios were rated as more arousing than happy scenarios in Experiment 1B. Self-relevance ratings consistently indicated that happy and neutral scenarios were perceived as more self-relevant than sad scenarios.

To test the feasibility of performing the pooled memory analyses, we examined if any significant differences emerged between experiments among the phenomenological ratings. Modeling the interaction of emotion category with experiment type (1A and 1B) revealed no significant main effects or interactions with experiment type when tested on ratings of valence (F_{2, 89.805} = 2.942, p = 0.058) or self-relevance (F_{2, 89.473} = 0.143, p = 0.867). Arousal ratings exhibited an interaction effect (F_{2, 89.785} = 3.151, p = .048), although follow-up pairwise comparisons of the interaction revealed no significant differences in arousal between the same emotion categories from each experiment (all ps > .05). Importantly, we also observed no significant differences between experiments in the relative differences among emotion categories in arousal ratings (all ps > .05). Thus, participants in Experiment 1B produced nearly identical distributions of phenomenological ratings as participants in Experiment 1A, suggesting similar engagement with the stimuli despite having imagined a different set of scenarios and in a different context (i.e., via an online task and not in the laboratory).

Mood Induction

Ratings of mood were acquired at four separate timepoints throughout the study: before encoding (T1), after encoding/before the mood induction (T2), after the mood induction (T3), and before the retrieval task on the next day (T4). We observed a significant two-way interaction of time and induction group when separately assessed in each experiment (Figure 2; Exp. 1A: F_2.42,188.44 = 29.734, p < .001; Exp. 1B: F_2.61,271.56 = 39.478, p < .001), which we further examined with pairwise comparisons between induction groups at each time point (see Table S2 for all comparisons).

Figure 2. Self-reported mood ratings in Experiments 1A and 1B.

Note. Ratings are grouped by experiment (1A and 1B) and post-encoding mood induction group (sad in blue, happy in green). The top row shows mood scores at each timepoint of the experiment with differences between groups only observed after the induction. Estimated marginal means for each timepoint are plotted with 95% confidence intervals. Mood scores were standardized within experiment since different mood questionnaires were used. More positive values indicate more positive mood. The bottom row shows density plots for post-encoding mood change. See Figure S11 for individual data points. ***p < .001.

For both experiments, our analysis revealed significant differences between groups only after the induction. We also examined consecutive differences between each timepoint within each group, which confirmed that, for both experiments, the happy and sad inductions resulted in a significant change in mood from T2 to T3. In Experiment 1A, but not 1B, both groups also reported significant changes in mood from before to after the encoding task (T1 to T2), although these effects are less critical as they occurred before the induction and were reflective of more neutral moods in both groups after participants had completed encoding. Importantly, when we tested the interaction of experiment (1A or 1B) with timing and mood induction group, we did not observe a three-way interaction that would indicate significant differences in mood ratings and induction strength between the two experiments (F_2.56,465.93 = 0.371, p = 0.742).

Retrieval

Post-Encoding Mood Change Predicts Mood-Congruent Biases in Next-Day Remembered Valence.

We first tested if mood during consolidation is associated with mood-congruent biases in how participants remembered the affective tone (valence) of the scenarios. As described in Methods, the model with the best fit to the data allowed for all possible interactions between encoding-valence, encoding-arousal, post-encoding mood change, and recall confidence. ANOVA testing of this model revealed a significant four-way interaction (with quadratic valence: F_1,7131.4 = 7.146, p = .008; with linear valence: F_1,6929.5 = 1.831, p = .176; see Table S3 for all fixed effects), further confirming that all four predictors were meaningful for predicting remembered valence.

To unpack the observed four-way interaction, we separately examined the three-way interaction of encoding-valence, encoding-arousal, and post-encoding mood change when model-derived, estimated marginal effects were averaged across levels of lower (−3 to 0) and higher (0 to 2) confidence (based on within-subject standardized values and their distribution range; average valence confidence at retrieval = 5.11, SD = 1.54; median valence confidence = 5). This analysis revealed a significant interaction of encoding-valence and encoding-arousal ratings with the slope of post-encoding mood change only at lower confidence (F_2,6532.36 = 9.781, p < .001), but not at higher confidence (F_2,5246.45 = 0.192, p = .825). We also did not observe a main effect of mood or other interactions with mood at higher confidence (all ps > .05)

We then further inspected the interaction of encoding-valence and post-encoding mood for less confident memories, when examined across lower (1–5) and higher (5–7) estimated marginal means for encoding-arousal (Figure 3; average encoding-arousal = 4.34, SD = 1.72; median encoding-arousal = 5; note that arousal values were standardized across participants in analyses to preserve scale interpretability). At lower arousal, scenarios that were originally rated as sad at encoding were recalled as being happier if post-encoding mood had been positive, but appropriately remembered as sad if post-encoding mood had been negative. In contrast, at higher arousal, neutral and moderately happy scenarios were remembered as happier if post-encoding mood change had been positive, but sadder if post-encoding mood had been negative.

Figure 3. Three-way interaction of post-encoding mood change, encoding-valence, and encoding-arousal predicting remembered valence (for less confident recall).

Note. All confidence bands and error bars represent 95% confidence intervals. The top row depicts model-derived estimated marginal trends and means averaged across levels of lower encoding-arousal (ratings of 1–5), while the bottom row depicts estimated marginal trends and means averaged across levels of higher encoding-arousal (ratings of 5–7). The left plots depict the relationship of encoding-valence with the slope of post-encoding mood change (T3 – T2) on remembered valence. Colors are assigned based on a gradient scale from sad (blue) to neutral (gray) to happy (green) valence ratings. Confidence bands that do not overlap with the red line at zero are significant at alpha = .05. Accordingly, significant regions are shown with full color, while non-significant regions are transparent. The right plots depict simple slopes for the relationship between post-encoding mood change and remembered valence at different levels of encoding-valence for sad (−1 to −0.5 SD valence), neutral (mean valence), and happy (+0.5 to +1 SD valence) scenarios. That is, each simple slope plot is drawn from a specific point on the x-axis in the leftmost plot, representing the original valence of the scenarios at encoding. See Figure S8 for a depiction of raw data values. ***p < .001, *p < .05.

In summary, mood-congruent biases in valence recall emerged only for low-confident memory responses, as highly-confident memories were generally more accurate and less prone to bias. The mood-congruent biases we observed were dependent on both the valence and arousal ratings of the scenarios at encoding, with lower-arousal sad scenarios and higher-arousal neutral/happy scenarios most representative of MCM. We note that higher-arousal neutral scenarios tended to be emotionally ambiguous or uncertain, which likely allowed for the valence bias to emerge relative to lower-arousal, unambiguously-neutral scenarios.

To evaluate the stability of these effects across experiments, we tested the same model but with Experiment now allowed to interact with all other terms. The observed four-way interaction remained significant and grew slightly stronger (quadratic valence: F_1,7107.6 = 10.521, p = .001; linear valence: F_1,7002.5 = 4.197, p = .041), even when accounting for all possible interactions with Experiment. However, there was a moderating influence of Experiment on this interaction, specifically for the nonlinear (quadratic) effect of valence (F_1,7107.6 = 6.655, p = .01). As depicted in Figure S7, the MCM we observed for lower arousal sad stimuli remained consistent across experiments. However, MCM at higher arousal shifted depending on experiment, due to a quadratic effect in Experiment 1A but a linear effect in Experiment 1B. Specifically, Experiment 1A primarily showed mood-congruent biases for neutral stimuli, whereas Experiment 1B showed mood-congruent biases for happy stimuli. While we are cautious not to overstate differences between experiments for these higher-level interactions due to reduced power and a considerable increase in model complexity when allowing for all interactions with Experiment, these findings do suggest subtle shifts in remembered valence for high arousal stimuli depending on experimental context.

Post-Encoding Mood Change Predicts Mood-Congruent Biases in Next-Day Recognition Accuracy.

In addition to recalling the emotional experience of each imagined scenario, participants were also asked to identify the critical word that filled in the blank. Participants generally performed well on this task, with average accuracy of picking the correct word at 81.14% (SD = 10.93%). Average outcome accuracy (picking the correct word or same-category lure) was at 91.64% (SD = 8.17%), indicating that even when participants were wrong it was usually because they remembered the general emotional outcome of the scenario but picked the wrong word. Accordingly, confidence on the recognition task was also higher than we observed at recall, with average recognition confidence across all responses at 5.66 (SD = 1.69; median = 6).

We hypothesized that accuracy would be best for scenarios that were emotionally congruent with a participant’s post-encoding mood change, but that these effects may also be moderated by other stimulus properties. Indeed, as described in Methods, the best fitting model included all possible interactions between encoding-valence, encoding-relevance, post-encoding mood change, and recall confidence. ANOVA testing of this model revealed a significant four-way interaction (z = 2.235, p = .025; see Table S4 for all fixed effects).

As with the recall analysis, we unpacked the observed four-way interaction for the recognition analysis by separately examining the three-way interaction of encoding-valence, encoding-relevance, and post-encoding mood change when estimated marginal effects were averaged across levels of lower (−3 to 0) and higher (0 to 1.5) confidence (based on within-subject standardized values and their distribution range). This approach revealed a significant interaction of valence and self-relevance with the slope of post-encoding mood change at higher confidence (z = 10.254, p = .001), but not at lower confidence (z = 0.002, p = .968). We also did not observe any other main effects or interactions with mood at lower confidence (all ps > .05). In other words, the relationship between post-encoding mood change and recognition accuracy shifted as a function of encoding-valence and self-relevance only for more confident responses. Note that this effect is opposite to what we observed with remembered valence, where lower confidence responses were more susceptible to bias. We suspect that this shift is related to a difference in the distribution of confidence ratings for recognition and recall responses. That is, because participants generally performed well on the recognition task (especially for remembering the general emotional outcome by picking the correct word or lure) and confidence ratings were skewed towards higher confidence, lower confidence on the recognition task was more indicative of random guessing.

We next examined the two-way interaction of encoding-valence and post-encoding mood change when evaluated across lower (1–5) and higher (5–7) estimated marginal means for self-relevance (Figure 4). For highly self-relevant scenarios, post-encoding mood change was positively associated with recognition accuracy for happy scenarios, but negatively associated with recognition accuracy for sad scenarios. That is, participants were more accurate in identifying the correct critical word for happy scenarios if their post-encoding mood had become more positive, but more accurate in identifying the correct critical for sad scenarios if their post-encoding mood had become more negative. Post-encoding mood change was unassociated with recognition accuracy at all levels of encoding-valence at lower self-relevance.

Figure 4. Three-way interaction of post-encoding mood change, encoding-valence, and encoding-relevance predicting recognition accuracy (for more confident recognition).

Note. All confidence bands and error bars represent 95% confidence intervals. The top row depicts model-derived estimated marginal trends and means averaged across levels of lower encoding-relevance (ratings of 1–5), while the bottom row depicts estimated marginal trends and means averaged across levels of higher encoding-relevance (ratings of 5–7). The left plots depict the relationship of encoding-valence with the slope of post-encoding mood change (T3 – T2) on recognition accuracy (log odds ratio). Colors are assigned based on a gradient scale from sad (blue) to neutral (gray) to happy (green) valence ratings. Confidence bands that do not overlap with the red line at zero are significant at alpha = .05. Accordingly, significant regions are shown with full color, while non-significant regions are transparent. The right plots depict simple slopes for the relationship between post-encoding mood change and recognition accuracy at different levels of encoding-valence for sad (−1.5 to −1 SD valence), neutral (mean valence), and happy (+1 to +1.5 SD valence) scenarios. That is, each simple slope plot is drawn from a specific point on the x-axis in the leftmost plot, representing the original valence of the scenarios at encoding. These simple slope plots have been back-transformed to depict recognition accuracy as the probability of picking the correct outcome (significance tests were performed on the link scale from the generalized linear model). See Figure S10 for a depiction of raw data values. **p < .01, *p < .05.

To evaluate the stability of these effects across experiments, we tested the same model but with Experiment allowed to interact with all other terms. The observed four-way interaction remained significant and grew slightly stronger (z = 2.378, p = .017) even when accounting for all possible interactions with Experiment. Moreover, the observed four-way interaction did not significantly differ by experiment (z = −1.604, p = 0.109; see Figure S9 for a visualization of effects split by experiment).

Interim Discussion

The results from Experiment 1 provide initial evidence supporting the hypothesis that emotional memory biases emerge with post-encoding mood change, even when assessed across different contexts, stimuli, and induction methods. However, MCM was highly sensitive to a specific combination of features. That is, our participants not only varied in their construction of the imaginative scenarios, but memory for these scenarios also varied in susceptibility to mood-related biases depending on their original phenomenological attributes, as well as confidence at retrieval. First, recalling the general emotional tone (i.e., remembered valence) of a previous event when only partially cued resulted in mood-congruent biases for lower-arousal sad scenarios and for higher-arousal neutral/happy scenarios. Second, recognition accuracy also shifted based on post-encoding mood change, with better memory observed for highly self-relevant, mood-congruent scenarios than mood-incongruent scenarios. Thus, we observed emotional memory biases both in how participants remembered the imagined episodic events (remembered valence), as well as precisely what occurred (recognition accuracy). Our results did reveal some variation in the expression of MCM between studies, specifically for the remembered valence of higher arousal scenarios. Nevertheless, a similar pattern of MCM emerged across the two experiments, and performing our analyses on the pooled dataset helped target generalizable effects that transcend individual study contexts.

When taken together, the results from Experiment 1 suggest that mood change during early consolidation produces distinct effects in subsequent emotional memory depending on the type of mood experienced. Increased sad mood after encoding reinforces the negativity of lower- arousal sad memories while also shifting ambiguous (neutral) information to be misremembered as more negative, and maintaining a more accurate representation of personal episodic events that were sad. In contrast, increased happy mood after encoding attenuates negative valence associated with lower-arousal sad memories, shifts neutral or happy information to be misremembered as more positive, and maintains a more accurate representation of personal episodic events that were happy.

To our knowledge, these findings are the first to show that post-encoding moods may also retroactively shape memory, as has been documented for fear, reward, novelty, and stress manipulations, albeit via potentially different mechanisms (Dunsmoor et al., 2022). Importantly, our results suggest that stimuli may not have to be weakly encoded to still be influenced by a durable change in mood taking place during consolidation. This finding converges with recent evidence demonstrating that post-encoding stress preserves memory for central negative objects while decreasing memory for peripheral, neutral backgrounds that were paired with those objects at encoding (Cunningham et al., 2021). That is, increased cortisol during consolidation facilitated preservation of the most salient aspects from encoding, and this was observed when using cortisol as a continuous measure, irrespective of whether participants were assigned to the control or stress groups. In a similar vein, our study suggests that the degree of mood change during consolidation also selectively targets specific information at encoding to be preserved in memory, but that this selectivity is associated with multiple attributes of a stimulus that may have already assigned priority at encoding (e.g., valence, arousal, and self-relevance). These attributes also define the congruency of the stimulus with the post-encoding mood, indicative of an associative network linking moods to specific emotional material (Bower, 1981). Retroactive memory enhancements may be overlooked in other studies if analyses do not incorporate subjective experiences of participants during encoding, whether in the context of mood or other post-encoding manipulations that have produced inconsistent findings.

Note, however, that MCM for remembered valence could also be attributable to source misattribution. That is, when less confident in a memory, participants may have conflated their experience imagining a scenario with their mood after encoding, rather than a change in how the stimulus was originally consolidated. However, given that post-encoding mood change did not seem to have a universal effect on memory (as indicated by the observed interactions with arousal), and that recognition accuracy for the exact word was also affected by post-encoding mood (despite overall good recognition accuracy and high confidence), our collective findings suggest that the observed effects cannot be attributed to source memory error alone. Nevertheless, more work is needed utilizing standard recognition tasks that can take into account how well participants remember previously-encoded content while also correctly identifying new content, even if that new content is mood-congruent. Thus, we designed our next experiment to address this limitation.

Experiment 2

Across multiple levels of analysis, the findings from Experiment 1 indicate the presence of MCM resulting from post-encoding mood. Yet, it remains unclear how specific such memory biases were to the stimuli and retrieval tasks that we used, which encouraged participants to construct and then reconstruct personal and imaginative scenes. Moreover, although we did find evidence for MCM in the recognition task, overall performance was high, and we were limited to assessing hit rates since participants only viewed the scenarios they had seen during encoding. We were thus unable to examine MCM in relation to memory discriminability, which involves measuring how well participants detect previously seen items while accounting for bias in their response patterns to new items. An individual may have both a high hit rate and high false alarm rate, for instance, indicating an inability to distinguish old and new items in memory. We hypothesized that if post-encoding mood strengthens the consolidation of mood-congruent material, then it would not only increase accuracy for mood-congruent stimuli as shown in Experiment 1, but also strengthen memory discriminability between old and new mood-congruent information.

To test this hypothesis, we redesigned our encoding task to use emotional images instead of emotional text scenarios. At retrieval, we then showed participants these same images during a surprise memory test, intermixed with new images that were not seen at encoding. Our design was informed by Kark and Kensinger (2019), who also used an old/new paradigm with images to examine retrieval biases associated with post-encoding functional connectivity in the brain. These authors showed that next-day biases in preferential recognition accuracy for positive or negative stimuli were associated with changes in amygdala-cortical coupling in the brain after the encoding task. Although the authors did not apply a post-encoding mood induction (or any manipulation, other than a resting-state brain scan), affective states after encoding may still have played a role in the emotional memory biases that were observed the following day.

Materials and Methods

Participants

All participants were recruited via the Prolific crowdsourcing platform in January 2023 and received monetary compensation for their time ($10 per session). Inclusion criteria included an age range of 18–39 years old, as well as no known history of a neurological condition or psychiatric disorder. All participants completed written informed consent in accordance with the Duke University Institutional Review Board guidelines. As with Experiment 1, our sample size was informed by a power analysis that indicated 120 participants are needed to detect a within-between interaction (e.g., Condition x Group effect) with 80% power when assuming a small/medium effect size of η_p² = 0.04 (using the as in SPSS effect size specification). Given that our encoding and retrieval sessions were separated by two days (to increase the difficulty of the memory task) and the study was completed online, we accounted for potentially high attrition rates by collecting data from more than 200 participants. In total, 213 participants were paid for completing the encoding session, while 159 participants were paid for returning and completing the retrieval session. Of those, nine participants were excluded for indicating technical issues at retrieval or that they were honestly not paying attention during the experimental tasks. Thereafter, we further excluded nine participants due to outlier behavior (> 3 SD) reflecting low accuracy on the active fixation task at the encoding or retrieval sessions (n = 6) or high levels of negative mood prior to the mood induction (n = 3). We also excluded one additional participant who straight-lined their valence and arousal ratings at encoding (i.e., the same ratings for all images). Applying these exclusions resulted in a final sample of 140 participants (61 women, 75 men, 4 non-binary; M_age = 29.7 yrs, SD = 5.6; 14 Hispanic or Latino, 126 not Hispanic or Latino; 5 Asian, 29 Black, 98 White, 6 more than one race, and 2 other), with 72 participants randomly assigned to the happy mood group and 68 participants to the sad mood group.

General Procedure

The general flow of procedures for this study (see Figure 5) was identical to that of Experiment 1B, with the notable exception of the structure of the encoding and retrieval tasks due to using images as stimuli instead of scenarios, as well as a longer delay between encoding and retrieval (two days instead of one to make the memory task more difficult). Participants completed an encoding task with images as stimuli (see the Encoding Task section for more details), and a retrieval task that consisted of identifying images as old or new (see the Retrieval Task section for more details). On average, the retrieval session was started by participants 52 hours (SD = 4.4) after the end of the first session. Additional procedural details are provided in Supplemental Materials.

Figure 5. Experimental design for Experiment 2.

Note. Mood ratings were acquired before encoding, after encoding/before the induction, after the induction, and before retrieval.

Stimuli

Images were selected from the International Affective Picture Systems (IAPS), using normative data provided by Lang et al. (2008) for valence and arousal ratings, as well as normative data provided by Libkuman et al. (2007) for discrete ratings of happiness and sadness. 132 images were selected, with 44 images assigned to each normative category. Importantly, we selected negative images that were sadder than they were angry, fearful, or disgusting. Normative ratings indicated that all three categories differed in valence (all ps < .001), while sad and happy images were equated on arousal (p = 1) but both more arousing than neutral images (both ps < .001). See Supplemental Materials (Figure S12) for an overview of all normative data. During encoding, 22 images per normative category were randomly selected for each participant. During retrieval, all 132 images were shown for all participants.

Degraded line-drawing versions of each image (similar to those used in Kark & Kensinger, 2019) were created using an in-house python script. We created these versions so that participants could be presented with partial cues during the recognition task (similar to Experiment 1), while also minimizing the presence of visual and emotion induction confounds during retrieval. For each image, we first used the pencilSketch function from OpenCV to generate a sketched version of the image, followed by decreasing and then increasing the contrast to emphasize object borders and diminish background noise. The final line-drawing version was created using the median_filter function from the SciPy package to further remove remaining noise.

Encoding Task

The encoding task consisted of viewing images and providing ratings of valence (very sad to very happy), arousal (very calm to very intense), and self-relevance of the depicted scenes (very low to very high) using keys 1–7 on the keyboard. For each participant, 22 images per normative category were randomly selected to be shown at encoding. On each trial, participants viewed a line-drawing sketch of the image for three seconds, followed by the full image for another four seconds. Participants were first shown the line-drawing sketch so that they were familiar with these versions of the image, given that only the line-drawing sketches were shown at retrieval. Ratings were then presented in a random order, each displayed for four seconds. Trials were separated by the same active fixation odd/even task as Experiments 1A and 1B. Trials were pseudorandomized so that each set of three trials consisted of one happy, one neutral, and one sad image in a random order.

Mood Induction

The mood induction task was identical to the procedures in Experiment 1B. Participants were first asked to provide a sad or happy memory and complete questions on episodic details for that memory while listening to mood-congruent music. They then watched a series of five sad or five happy news clips, followed by a sad or happy film clip depending on group assignment.

Retrieval Task

During the retrieval task, all participants viewed all 132 images. Half of these images were shown at encoding (randomly selected for each participant), and half of these images were new. For each trial, the line-drawing version of the image was displayed for a total of 10 seconds. For the first two seconds, participants viewed only the image on the screen, and for the remaining 8 seconds they had to decide if the image was old or new and provide a confidence rating for their response (1–7; not at all confident to very confident). Trials were separated by an active fixation odd/even task. Like Experiments 1A and 1B, participants provided responses at retrieval by clicking on the screen instead of using the keyboard.

Statistical Analyses

All statistical analyses were performed in R. ANOVAs (Type III sum squares) of encoding ratings, mood induction response, and retrieval performance were computed with the afex package (version 1.3.0; Singmann et al., 2023), and follow-up comparisons with the emmeans package (version 1.8.7; (Lenth, 2016; Lenth et al., 2023). We used Greenhouse-Geisser sphericity correction for degrees of freedom, and follow-up comparisons are reported after Bonferroni correction for multiple comparisons.

For the retrieval analysis, we did not use the same mixed-effects modeling approach as Experiment 1 for two primary reasons. First, participants in Experiment 2 only provided ratings for old images that were seen at encoding, and thus we were unable to incorporate item-specific responses for all images across all participants. That is, we were unable to relate encoding ratings to the likelihood of participants having false alarms or correct rejections towards new items, as these items were never seen at encoding. Second, Experiment 2 was specifically designed to evaluate memory sensitivity by comparing, within each subject, the ability to appropriately recognize previously seen images as old (i.e., a high hit rate) with the ability to appropriately reject new images (i.e., a low false alarm rate). Whereas the forced-choice recognition task used in Experiment 1 inherently controls for response bias, the old/new paradigm used in Experiment 2 is susceptible to response tendencies that must be accounted for (e.g., a tendency for a participant to respond “old” to most items). To this end, we computed aggregated memory performance scores for each participant that integrate hit rate, false alarm rate, and confidence ratings, thereby providing a more robust measure of memory sensitivity than trial-level accuracy.

Various methods are used in the memory literature to analyze old/new memory paradigms, including the use of corrected hit rates or d’ metrics. However, these metrics do not directly incorporate the strength of memory signals into analyses (e.g., confidence ratings), instead making assumptions of the distribution of latent memory signals that may be incorrect (Brady et al., 2022). An alternative approach that is more valid in representing memory sensitivity (and therefore suggested to be used with old/new recognition performance) is receiver operating characteristic (ROC) analysis (Brady et al., 2022).

Following this recommendation, for each participant, we first computed their hit and false alarm rates at the highest criterion (a confidence rating of 7) by dividing the number of hits/false alarms at this level by the total number of genuinely old/new items in a given category of stimuli. We then examined the same rates when computed at a more liberal criterion (confidence ratings of 6–7). We continued calculating the proportion of hit and false alarm rates until reaching the most liberal criterion (the full range of available confidence ratings, 1–7). The calculated series of hit and false alarm rates can then be plotted as an empirical ROC curve, with the curve ending at (1,1). By generating these metrics, we were able to assess how the relationship between hit rate and false alarm rate changed at different response criteria. A participant responding at random will show a perfect linear relationship between hit and false alarm rate, whereas the ROC curve becomes more curvilinear as hit rates outpace false alarm rates. Importantly, this ROC approach to analyzing memory recognition data allowed us to appreciate asymmetric shapes of the ROC curve by computing the full spectrum of responses across levels of memory strength (Brady et al., 2022). For each participant and emotion condition, we computed ROC curves in this manner, and then calculated the area under the curve (AUC) using the AUC function from the DescTools package (version 0.99.49; Signorell, 2023). See Figure S14 for example ROC curves for individual participants.

Due to our specific interest in MCM, our primary analyses focused on directly contrasting happy and sad recognition performance. This was achieved either by computing the difference in AUC scores between happy and sad conditions in order to represent emotional memory bias (similar to Kark and Kensinger, 2019), or only including emotional stimuli in ANOVA models that further tested the moderation of arousal or self-relevance levels. Neutral stimuli were excluded from these models given that, unlike in Experiment 1, there was less variability in these ratings and we were unable to incorporate continuous, trial-level data to model graded changes in valence, arousal, or self-relevance strength with our moderation analyses.

Transparency and Openness

We report how we determined our sample size, all data exclusions, all manipulations, and all measures in the study. The de-identified data and analysis code on which the results and conclusions are based are available at the following Open Science Framework website (Faul & LaBar, 2025): https://osf.io/m3nz7/. This design and its analysis were not pre-registered.

Results

Encoding

An overview of all image ratings acquired during encoding and the corresponding statistical report is provided in Supplemental Materials (see also Figure S13 for a visual display). In brief, we observed two primary differences compared to the normative data. First, valence ratings for six of the sad images were rated more neutral by our participants than the normative data had suggested. We accounted for this discrepancy in subsequent memory analyses by reassigning these images to the neutral condition. Second, sad images were rated as significantly more arousing than happy images by our participants, which may confound the presence of mood-related biases in memory. To help account for this confound in memory analyses, we separated the data into higher and lower arousal levels via a median split within each emotion category. In doing so, the lower arousal sad images were comparable in arousal ratings to the higher arousal happy images (p = .099). We also applied this same approach to the self-relevance ratings. For subsequent memory analyses, we examined the original groupings of stimuli into happy and sad sets, as well as these subgroupings into low/high arousal and self-relevance sets.

Mood Induction

Mood ratings exhibited a significant two-way interaction of time and induction group (Figure 6, F_2.31,318.88 = 50.67, p < .001), which we further examined with pairwise comparisons between groups at each time point (see Table S5). Replicating the effects seen in Experiments 1A and 1B, we again observed significant differences between groups only after the induction. For both groups, we also observed a significant change in mood from before to after encoding (to a more neutral mood), as well as a significant change in mood from before to after the induction.

Figure 6. Self-reported mood ratings in Experiment 2.

Note. Ratings are grouped by post-encoding mood induction group (sad in blue, happy in green). The left plot shows mood scores at each timepoint of the experiment with differences between groups only observed after the induction. Estimated marginal means for each timepoint are plotted with 95% confidence intervals. More positive values indicate more positive mood. The right plot shows a density plot for the distribution of post-encoding mood change. See Figure S11 for individual data points. ***p < .001.

Retrieval

ANCOVA testing of recognition performance across all available images (with post-encoding mood change as a continuous covariate) revealed a significant main effect of emotion category (F_1.96,270.03 = 7.369, p < .001), such that both happy and neutral images were generally remembered better than sad images (H-S: b = 0.028, t₁₃₉ = 3.659, p = .001, 95% CI = [0.009, 0.046]; N-S: b = 0.026, t₁₃₉ = 2.966, p = .011, 95% CI = [0.005, 0.046]; see Figure S15). The main effect of mood change was not significant (F_1,138 = 0.788, p = 0.376), nor was the interaction of mood with emotion category (F_1.96,270.03 = 0.952, p = 0.386). To specifically examine MCM, we also evaluated the relationship between post-encoding mood change and emotional memory bias scores, calculated as the difference between happy and sad recognition performance. While post-encoding mood change was positively associated with better performance for happy versus sad recognition, this effect did not reach statistical significance (r = 0.123, p = 0.148).

We next tested whether the relationship between post-encoding mood and emotional memory bias was moderated by the arousal or self-relevance of the images at encoding. To examine all possible pairings, we separately coded each category of emotional stimuli prior to computing AUC scores (e.g., happy low arousal, happy high arousal, sad low arousal, and sad high arousal) and then tested the interaction of category assignment with post-encoding mood change. For arousal ratings, we observed a significant two-way interaction of post-encoding mood change and emotion category (F_2.98,410.81 = 3.594, p = .014). Examining all possible comparisons between categories (corrected for multiple comparisons) only revealed a significant difference in the mood-memory relationship for higher arousal happy images and lower arousal sad images (higher arousal happy – lower arousal sad: b = 0.035, t₁₃₈ = 3.13, p = .013, 95% CI = [0.005, 0.064]; Figure 7B). In other words, more positive post-encoding mood was associated with a stronger difference in recognition performance for higher arousal happy images and lower arousal sad images. Given that high arousal happy images were comparable in arousal to low arousal sad images (see Supplemental Materials, Figure S13), this effect indicates that emotional memory bias scores were most affected by post-encoding mood change when arousal was equated across the sad and happy image sets. We did not observe the same interaction when emotional images were grouped based on self-relevance (F_2.95,407.36 = 1.44, p = .231).

Figure 7. Relationship of post-encoding mood change with emotional memory bias scores.

Note. Effects are shown when assessed across all happy and sad images (A), when equated on arousal (B), and across the most personally relevant stimuli (C).

For each panel, the top plot shows the slope of post-encoding mood change (T3 – T2) on emotional memory bias when examined across all subjects versus those with more reliable memory performance (total AUC > 0.75). The bottom plot shows a scatter plot and regression line for the relationship between post-encoding mood change and emotional memory bias scores (only for participants with reliable performance). Positive values indicate better memory for happy images, while negative values indicate better memory for sad images. Each dot represents a different participant, color coded based on induction group (sad in blue, happy in green). All of the depicted effects are extracted from models including overall performance as a between-subject variable. **p < .01; *p < .05.

In exploratory post-hoc analyses, we tested whether the presence of emotional memory bias was further impacted by participants’ overall discriminative ability. That is, we wondered whether participants with low AUC scores yield uninterpretable or unreliable measures of memory bias, given that lower values approach chance-level performance (i.e., random guessing). We therefore used total AUC scores calculated across all images to separate participants into groups with reliable (AUC > 0.75, n = 93) and unreliable (AUC < 0.75, n = 47) overall discriminability, when based on thresholds used in the literature to assess diagnostic performance (Fan et al., 2006). Adding this grouping variable into our models did not further moderate the significant two-way interaction of post-encoding mood change and emotion categories split on arousal (F_2.98,404.71 = 1.52, p = .208), suggesting a consistent effect across participants that remained significant even when accounting for overall performance (F_2.98,404.71 = 3.324, p = .02). However, overall performance did significantly moderate the effect of post-encoding mood when assessed across all images (F_1,136 = 4.597, p = .034), such that only participants with reliable performance exhibited a significant relationship between post-encoding mood change and emotional memory bias (b = .026, t₁₃₆ = 2.566, p = .011, 95% CI = [0.006, 0.046]; Figure 7A). Participants’ overall performance also moderated the effect of self-relevance, as evidenced by a significant three-way interaction of performance, self-relevance, and post-encoding mood change (F_2.96,402.43 = 3.150, p = .026). Examining all possible pairwise comparisons within each performance group (corrected for multiple comparisons within group) indicated a significant difference in the relationship of mood change with memory performance only when comparing highly self-relevant happy and sad images in participants with overall reliable performance (happy higher relevance – sad higher relevance: b = .045, t₁₃₆ = 3.130, p = .013, 95% CI = [0.006, 0.083]). In other words, for participants with reliable recognition data, post-encoding mood was associated with a stronger emotional memory bias for mood-congruent images that were higher in self-relevance (Figure 7C). No other differences between self-relevance categories were observed (all ps > .05)

In summary, analysis of memory performance -- when examined across all available images and within subgroups of stimuli -- indicated that MCM was primarily evident across all participants when happy and sad images were equated on arousal. However, exploratory post-hoc analyses further revealed that MCM was sensitive to overall memory discriminability, as participants with more reliable recognition performance showed MCM even when assessed across all images, and to a greater degree for images with higher self-relevance. These findings suggest that when memory recognition is particularly poor, the effects of post-encoding mood are overshadowed and thus less evident. For completeness, we have depicted all of these effects in Figure 7.

Interim Discussion

Our results indicate that MCM resulting from a post-encoding mood shift can be shown with other stimuli beyond imaginative scenarios, as well as with a different memory test. However, like Experiment 1, we again demonstrate that the emergence of such biases is moderated by several key variables. Following recent recommendations (Brady et al., 2022), we incorporated confidence ratings into the memory analyses by generating ROC curves for each participant that were sensitive to the relationship of hit and false alarm rates at different levels of latent memory strength. Our analyses revealed that the strongest MCM biases emerged when higher-arousal happy images were compared to lower-arousal sad images. This finding is in line with those from Experiment 1 that showed lower-arousal sad stimuli and higher-arousal neutral/happy stimuli were more susceptible to mood-related effects, although note here that the “low” arousal sad images and “high” arousal happy images were effectively equated on arousal. In exploratory analyses, we further found that MCM was more evident among participants with reliable discriminatory performance, suggesting that particularly poor recognition accuracy may obscure any mood-related biases. Thus, the incorporation of overall discriminatory performance into our models effectively reduced the influence of noisy, idiosyncratic subjects.

The findings from Experiment 2 further support the notion that post-encoding moods preserve memory for mood-congruent content. Participants were best at correctly identifying previously seen content congruent with their post-encoding mood, while also correctly rejecting new content with a similar valence. That is, our participants were not biased towards identifying any item as old simply because it matched in valence with the mood they had experienced two days prior. Rather, discriminative ability for mood congruent content was prioritized, and again these effects were sensitive to phenomenological attributes of the stimuli at encoding that made them more or less sensitive to the effects of post-encoding mood.

General Discussion

Decades of research has shown that moods selectively bias memory towards mood-congruent content (Faul & LaBar, 2023). However, these effects were almost exclusively studied in the laboratory by inducing moods immediately prior to encoding or retrieval and were often only examined for immediate memory tests. Until now, it was unclear whether and how MCM effects emerge from moods that are induced after encoding during the initial stages of memory consolidation. Here we have shown that a shift in mood during initial consolidation is associated with mood-congruent biases in memory for emotional appraisals as well as mood-congruent biases in recognition accuracy. Our findings advance an understanding of emotional memory bias and relevant moderators of MCM, while also demonstrating a novel approach to studying the effects of mood after encoding.

Importantly, though, MCM only emerged when specific conditions were met. Across both sets of experiments, our results revealed moderating roles for the self-relevance and arousal of stimuli in shaping the presence of MCM. We hypothesized that self-relevance would influence the extent of MCM based on existing literature that has shown the effects of mood are strongest for self-relevant material (Faul & LaBar, 2023), presumably due to such information being more closely integrated within an associative network of emotional content. In Experiment 1, recognition biases emerged only for highly self-relevant scenarios. Similarly, in Experiment 2, recognition biases were more present among happy and sad images that were higher in self-relevance. The importance of self-relevance in shaping MCM suggests that mood-emotion congruence may depend on more than just the valence of affect, but also the self-relevance of both the experienced mood and emotional stimulus. Indeed, we specifically incorporated self-referential elements (e.g., autobiographical memories) to all mood inductions to encourage the generation of self-referential moods. Although both emotional and self-referential information are prioritized in memory, their contributions are often studied separately. Recent empirical perspectives suggest that these two domains may rely on shared neural mechanisms that facilitate more elaborative encoding and increased rehearsal to collectively boost episodic memory (Gutchess & Kensinger, 2018). These shared mechanisms may explain why MCM is bolstered when overlap between moods and emotions is present for both valence and self-relevance.

Note, however, that we did not observe a role for self-relevance in shaping all aspects of MCM. In Experiment 1, recall performance for the valence of the partially-cued scenarios was instead moderated by stimulus arousal. The reduced influence of self-relevance may, in part, be due to our scenario task encouraging self-referential processing for all the imagined events, and thus we assessed memory for items that were already personally relevant by design (i.e., all scenarios referred to the participant in the second-person perspective). This generalized self-referential mode of encoding may have obscured the presence of item-specific self-referential effects. While we did not make specific a priori predictions regarding a differential influence of phenomenological attributes on recall or recognition tasks, our results suggest that recall performance for the remembered valence of scenarios may simply be less sensitive to self-referential moderation. It is important to note that the recall task assessed memory for the general gist of the scenario outcomes while the recognition task evaluated memory for specific details (critical words). Self-relevance may be less important for retaining the general gist of mood-congruent scenarios, but remains important for preserving specific details of mood-congruent memories. This interpretation is consistent with prior work showing a stronger influence of self-referencing on retaining specific, versus general, memory details (Serbun et al., 2011).

Across our experiments, we also observed moderating effects of stimulus arousal. We anticipated arousal may play a role in MCM given its documented influence in shaping emotional memory biases (Faul & LaBar, 2020; LaBar & Cabeza, 2006; McGaugh, 2004), and therefore considered the possibility that more arousing stimuli are prioritized during encoding and better integrated with a congruent mood in an associative network of affective information, thereby resulting in greater susceptibility to mood during consolidation. However, we did not predict a priori that the influence of arousal would primarily emerge for higher arousal neutral-happy stimuli, whereas sad mood during consolidation seemed to preferentially target lower arousal sad stimuli, suggesting a more nuanced relationship of arousal with MCM. Higher arousal may only boost the presence of MCM for affective experiences that typically evoke high arousal. That is, happiness is generally considered to be a high valence and high arousal affective state, whereas sadness is a low valence and low arousal affective state (Posner et al., 2005). Therefore, in the present study, post-encoding moods retroactively biased memory for emotional items that were prototypically similar in both valence and arousal to the induced moods, rather than universally acting upon high arousal stimuli. Indeed, as originally postulated in associative network theory, activated mood nodes will spread activation to linked representations with similar affect, including those with similar patterns of autonomic arousal and expressive behavior (Bower, 1981). These findings also converge with similar effects seen in the depression literature, where individuals with depression exhibit negative memory biases primarily for lower-arousal negative stimuli (Deldin et al., 2009) and generally recall more low arousal autobiographical memories (Young et al., 2012). Note, however, that in Experiment 2 the lower-arousal sad images and higher-arousal happy images were effectively equated on average arousal ratings. Future research on MCM for images may benefit from a stimulus set that covers a wider range of arousal for images specific to sadness and happiness than we were able to use here.

Taken together, the effects we observed suggest that post-encoding moods retroactively act upon self-relevant, valence- and arousal-congruent stimuli to bias subsequent memory. How exactly these effects emerge at a neural level remains to be discerned. The associative network theory of memory and emotion proposes that when mood nodes are activated, this activation spreads along established links to neighboring affective nodes (Bower, 1981). During consolidation, an activated mood will reactivate affectively congruent content and strengthen the storage of this information into memory. Yet, it remains unclear whether such mood nodes exist, as well as how to isolate a neural signature of spreading activation in humans. Some insight may be gained from recent neuroimaging work on retroactive memory enhancement using fear conditioning tasks. Clewett et al. (2022) examined the brain mechanisms that support enhanced memory for image categories that were selectively conditioned to be aversive after encoding. Across participants, greater retroactive memory enhancement effects were associated with a greater increase in functional connectivity between the hippocampus and category-selective cortical regions from before to after the conditioning task (Clewett et al., 2022). Extending to the present findings, individuals who exhibit stronger MCM from post-encoding mood may also display increases in hippocampal connectivity with emotion-selective cortical regions that strengthen the consolidation of affectively congruent content. Such work, however, requires identifying neural signatures of emotional states, if such discrete markers exist. Multivariate decoding techniques have shown promising results indicating that different emotional states can be reliably classified from fMRI data (Horikawa et al., 2020; Kragel & LaBar, 2014, 2016). Similar applications to MCM research would be a fruitful avenue for future researchers to explore. See Supplementary Materials for further discussion on neurocognitive mechanisms and, relatedly, the timing of post-encoding effects.

It is important to note that in all experiments, some participants did not respond to the mood induction or only minimally shifted in mood. These non-responders are prevalent across mood induction studies and are often removed from analyses (Rottenberg et al., 2018). However, such an approach necessitates deciding on a specific threshold for response, as well as which subjective mood measure(s) to use in this calculation. In our analyses, for instance, we analyzed aggregated mood scores across multiple subscales. If we had implemented a threshold for removing non-responders we could have done so with the aggregated scores or with specific scales (e.g., only for ratings of happiness and sadness). We avoided the ambiguity of this decision by instead modeling post-encoding mood change as a continuous variable in all analyses (rather than separating by group), thereby appreciating the full range of mood change across all participants and retaining non-responders. This approach was particularly advantageous because it suggests that the observed MCM effects are associated with the strength of mood during consolidation, as opposed to other confounding factors across the two induction types. For example, systematic differences in the content of autobiographical memories or video clips between the two groups may have reactivated specific items from encoding, not because of affective congruency but because of overlapping conceptual features. Yet, by modeling post-encoding mood change as a continuous measure, we specifically tested for MCM effects that scaled with the strength of the induction response rather than group assignment. Note that this approach is conceptually similar to Cunningham et al. (2021), who found that the degree of post-encoding endogenous cortisol concentrations shaped the strength of a next-day emotional memory trade-off effect. See Supplementary Materials for further discussion of how our induction materials supported the generation of diffuse, graded mood states across participants.

Limitations and Constraints on Generality

Some limitations should be considered with the current findings. Our results suggest that the influence of post-encoding mood on next-day memory is sensitive to multiple factors, including stimulus characteristics at encoding (i.e., valence, arousal, and self-relevance), the degree of mood change, and confidence judgements at retrieval. Future work is needed to validate the role of these moderators, as well as additional factors not tested here (e.g., the length of the encoding-retrieval delay). Our findings are intended as initial evidence for the presence of MCM resulting from post-encoding mood, but more work is needed to assess the reliability and generalizability of the interactions we have observed here.

We only manipulated happy and sad moods; the extent to which the findings generalize to other categories of emotional experience remain to be tested. Furthermore, as with other mood induction studies, concerns over demand characteristics are warranted and may have influenced the self-report ratings from participants and/or the focus of their attention. In many aspects of our design, however, we attempted to mitigate the potential influence of demand characteristics via participant blinding, validated inductions, and reliable mood measurements (for a detailed discussion, see Supplemental Materials). Moreover, although we attempted to isolate the effects of mood on memory during early consolidation, it is possible that our findings were influenced by other means, such as lingering moods at retrieval, or that participant’s memory for the mood they experienced after encoding was used as a general heuristic to guide their recall and/or recognition of the stimuli. It is important to note, however, that in all experiments the two groups of participants provided similar mood ratings at retrieval, and only differed in mood ratings after the induction. Critically, the effects we observed consistently demonstrated selective, mood-congruent biases that interacted with phenomenological attributes of the encoded stimuli, rather than generalized effects across stimuli.

Several factors further constrain the generality of our findings. The first day of our experiments ended with the post-encoding mood induction, and therefore we did not control or measure participant behavior after finishing the session. Certain regulatory strategies and/or trait dispositions may have systematically shaped post-encoding consolidation processes in ways we did not appreciate here. Moreover, these experiments were only performed with younger adults, yet mood effects may differ among older adults. Future work is also needed to determine whether our findings remain consistent across nationalities and cultures. Finally, although we used a continuous measure of mood change in our analyses, we did not incorporate a control/baseline (no induction) group into our design and were therefore unable to directly assess enhancing or impairing effects of mood on memory in relation to no induction after encoding. Note, however, that we did perform sensitivity analyses that controlled for the influence of other post-encoding changes in emotional states, including tension, anger, and fatigue, which reproduced the same findings reported here (see Supplemental Material).

Conclusions

In conclusion, the present findings provide rigorous, novel evidence that post-encoding moods can retroactively bias memory for emotional content in a mood-congruent direction. The study reorients prior MCM work by inducing mood during consolidation rather than prior to encoding or retrieval. Thus, even interim, offline moods generated from other emotional sources can have significant consequences on how previously-experienced emotional events are ultimately remembered. We adopted a systematic approach to our experimental designs that addressed gaps in the MCM literature while also adhering to known moderators of MCM. This approach involved not only inducing mood during consolidation, but also testing memory at least a day later, comprehensively norming all stimulus materials to ensure discrete affective responses, encouraging self-referential processing of the encoded stimuli, testing both recall and recognition performance, and inducing both happy and sad moods that were diffuse, non-specific, and self-referential.