Encoding in a social feedback context enhances and biases behavioral and electrophysiological correlates of long-term recognition memory

Encoding often occurs in social contexts, yet research has hardly addressed their role in verbal memory. In three experiments, we investigated the behavioral and neural effects of encoding context on memory for positive, negative, and neutral adjectives, contrasting a social-feedback group (N = 24) with an explicit verbal-learning (N = 24) and a levels-of-processing group (N = 24). Participants in the social-feedback group were not aware of a recognition session one week later, but their memory was better than the explicit learning or the levels-of-processing groups'. However, they also exhibited the strongest response bias, particularly for positive words. Brain event-related potentials (ERPs) revealed largest early negativities (EPN) and late positivities (LPP) in the social-feedback group. Only in the subsequent slow-wave did the explicit learning group show higher amplitudes than the other two groups, suggesting reliance on strategic rather than automatic processes. Still, context-driven incidental encoding outweighed explicit instructions, specifying a decisive role of social factors in memory.

Humans are a social species. Therefore, human memory formation, as much other cognitive activity, often occurs in social contexts. Yet, scientific memory research has paid limited attention to the social situatedness of memory encoding, taking memory operations devoid of contextual embedding as its point of departure 1 , thereby delineating many key principles. Meanwhile, contextual binding in space and time has proven crucial for successful episodic memory formation 2 . The social encoding context has, by comparison, received little attention, although recently, social influences on memory systems have attracted scientific interest. These range from investigations into brain mechanisms of social memory conformity 3 to the finding that the hippocampus, a critical region for spatial navigation and episodic memory formation, also represents a "social space" of relationships 4 . Such findings lend momentum to the hypothesis that the social context in which items are encoded affects their subsequent remembering.
Memory research also initially focused on content-general mechanisms. However, human memory differs depending on the to be encoded material and several content-related factors have been shown to affect episodic memory. For instance, ample evidence indicates that emotionally relevant material is often remembered differently and often better than neutral material 5,6 , which may be due to both higher accuracy and more pronounced response bias for emotional contents 7,8 . Moreover, such emotion effects may be driven by arousal, equally affecting positive and negative contents, or by positive or negative valence 5,9 . Next to emotional significance, the material's self-reference also plays an important role 10,11 . Whereas emotional significance is typically evaluated via normative ratings, self-reference is typically experimentally induced by a corresponding processing instruction on individually presented words, which is known to enhance recognition or recall. This "self-reference effect" is attributed to elaborative stimulus processing and deep encoding 12 , similar to other levels of processing manipulations, such as concreteness decisions 13 . Setting self-referential processing apart from other encoding manipulations, recent theories propose that the "self " acts as a hub, enabling Results Behavior during the encoding task. Social-feedback group (feedback acceptance). Positive feedback was more often accepted than negative and neutral feedback (ps < 0.001), while neutral feedback was also more often accepted by participants than negative feedback (p = 0.001) as reflected by a main effect of emotional content (F (2,46) = 62.99, p < 0.001, η p 2 = 0.733) and subsequent follow-up tests. A smaller main effect of feedback sender (F (1,23) = 5.12, p = 0.033, η p 2 = 0.182) revealed that participants were more likely to accept feedback from the more relevant ("human") sender. No interaction was found (F (2,46) = 0.48, p = 0.622, η p 2 = 0.020). Reaction times were 913 ms on average. While they did not differ significantly between sender conditions (F (1,19) = 4.09, p = 0.057, η p 2 = 0.177), they were descriptively longer for decisions on human feedback (999 ms) compared to computer feedback (827 ms). There were no significant differences between emotion conditions and no interaction (Fs < 0.81, ps > 0.452).

Recognition memory (one week later).
Recognition memory data are shown in Fig. 1 and further detailed in Table 1 and Supplement A.
Discrimination accuracy. As shown in Fig. 1, left panel, experimental groups differed in discrimination accuracy (F (2,65) = 7.57, p = 0.001, η p 2 = 0.189). The social-feedback group was significantly more accurate than both the verbal-learning (p = 0.007) and the levels-of-processing group (p < 0.001). The latter two groups did not differ (p = 0.403). Group and emotion did not interact (F (4,130) = 0.45, p = 0.776, η p 2 = 0.014), but a three-way interaction between group, encoding condition, and emotion (F (4,130) = 20.53, p < 0.001, η p 2 = 0.387) indicated that depending on the experiment, emotion effects differed between conditions (see Table 1). Within the social feedback group, discrimination accuracy did not differ between encoding during human compared to computer feedback for negative (p = 0.120), neutral (p = 0.330), or positive words (p = 0.774). Likewise, for the verbal-leraning group, no encoding differences between block A and block B were found for all three emotion conditions (ps > 0.441). Within the levels-of-processing group, encoding in the self-reference task compared to the concreteness task led to higher accuracy for positive words (p = 0.003) but lower accuracy for neutral words (p = 0.001). For negative words, accuracy did not differ between the self-referential and concreteness tasks (p = 0.113).

Response bias.
Response bias differed between the groups (F (2,65) = 10.31, p < 0.001, η p 2 = 0.241; see Fig. 1, right panel). Both the social-feedback and the verbal-learning group responded more liberally than the levels-of-processing group (p < 0.001 and p = 0.002, respectively). Group interacted with emotion (F (4,130) = 9.09, p < 0.001, η p 2 = 0.218; see Table 1), suggesting specific bias differences between the groups. We compared the  www.nature.com/scientificreports/ bias per emotion category, showing group differences for positive words (F (2,65) = 25.367, p < 0.001, η p 2 = 0.438), with a larger positive-bias in the social feedback group compared to the verbal-learning (p = 0.020) and levels-of-processing group (p < 0.001; see Fig. 1, right panel). There was also a more liberal responding to positive words in the verbal-learning than in the levels-of-processing group (p < 0.001). Response bias differed also for neutral words (F (2,65) = 7.03, p = 0.002, η p 2 = 0.178), showing no difference between the social feedback and verbal-learning group (p = 0.208), but more conservative responding in the levels-of-processing group compared to the other two groups (ps < 0.05). Finally, response bias group differences were also found for negative words (F (2,65) = 4.94, p = 0.010, η p 2 = 0.132), again showing no difference between the social feedback and verballearning group (p = 0.595), but more conservative responding in the levels-of-processing group compared to the other two groups (ps < 0.05). Finally, there was a three-way interaction between group, condition, and emotion concerning response bias (F (4,130) = 6.42, p = 0.001, η p 2 = 0.165). Follow-up investigations within groups revealed for the social-feedback no significant bias differences between encoding words in the human or computer block for positive (p = 0.882) and negative words (p = 0.053), but more liberal responding to neutral words encoded as human feedback (p = 0.032). For the verbal-learning group, bias did not differ between the three emotion conditions (ps > 0.288). For the levels-of-processing group, encoding words in the self-relevance compared to the concreteness task led to more liberal responding to positive (p = 0.004) and negative words (p = 0.019), but not for neutral words (p = 0.126).
Run 2/passive viewing re-presentation of word stimuli. After debriefing, a passive viewing run was designed to test for the immediate effects of the preceding experimental setup. There were no effects of the experimental group, emotion, or any interactions between these factors for the EPN, LPP, and the slow waves (Fs < 1.40, ps > 0.236).

Correlations between encoding ERPs and memory performance.
Correlations showed the strongest relationships between ERPs and memory indices for LPP amplitudes (see Table 2). For the social-feedback group, parietal LPP amplitudes were related negatively to the false alarm rate. Bayes factor analysis revealed moderate evidence of a relationship for positive words. Although the levels-of-processing group showed descriptively a similar relationship pattern over parietal sensors, Bayes Factors did not support such a relationship. For the verbal-learning group, the strongest relationships of LPP amplitudes were found over the frontal cluster.
Here, strong evidence was found for a relationship between ERP amplitudes during encoding and hit rate during recognition for neutral words. A similar pattern was found concerning negative and positive words, but Bayes Factors showed only anecdotal evidence of a relationship.

Discussion
Humans typically acquire information in social contexts. Therefore, we investigated its role for long-term recognition memory and encoding ERPs for positive, negative, and neutral words. We compared a group for whom a social context was induced via a self-introduction followed by evaluative feedback ("social-feedback" group) with a verbal-learning and a standard levels-of-processing group. Although the social-feedback group was not aware of the recognition session one week later, this group showed superior recognition accuracy, but also an increased response bias, particularly for positive feedback adjectives. Encoding ERPs revealed the highest EPN and LPP amplitudes in the social-feedback group. Only in the slow-wave window were ERPs most positive-going in the verbal-learning group. ERPs from the levels-of-processing group fell in between the other two. Together, these findings reveal a distinct effect of evaluative social context, enhancing long-term recognition memory and encoding ERPs beyond both explicit learning and standard levels of processing manipulations, while also inducing specific biases. The highest memory accuracy in the social feedback group indicates that the social context facilitated memory encoding, likely via elaborative stimulus processing which is known to enhance memory [e.g., see 11,12,31 ]. Since explicit learning tasks are typically easier and result in better memory than incidental ones 32,33 , these findings highlight the power of social context to increase stimulus encoding.
Groups also differed in their response bias, with less conservative responses in the social-feedback and the verbal learning groups. Moreover, the social-feedback group was specifically biased towards classifying positive words as "old". Healthy participants are known to have optimistic feedback expectations 34 , leading to distorted memory when information is incongruent with their positive self-view 35 . In fact, better than expected evaluations induce positively biased self-updating 36 as well as a specific feedback-related positivity 37 . Self-serving tendencies also occurred in feedback acceptance (social-feedback group) and self-descriptiveness decisions (levels-of-processing group), the highest acceptance rates for positive adjectives corresponding with positive self-views of healthy participants [e.g., see 34 ].
Concerning ERPs, words in the social feedback context elicited the largest EPN and LPP amplitudes, followed by levels-of-processing and, finally, instructed learning. Per se, these ERP modulations are broadly in line with findings of increased EPN and LPP amplitudes for various manipulations of self-reference and social feedback 19,38,39 . Given the EPN's role in early attentional selection 22 , results show a rapid prioritizing role of socialevaluative context beyond mere task-relevance. The highest LPP in the social feedback group is also broadly in line with the view that P3b-like components, such as the LPP, index successful memory encoding 25 www.nature.com/scientificreports/  www.nature.com/scientificreports/ appraisal were subsequently better remembered. Such items also elicited a larger fronto-central subsequent memory positivity than neutral ones. Whereas we have no subsequent memory ERP data for the present set of experiments, the much larger encoding positivities we found in the social context condition in parallel with the considerably higher recognition rates are certainly consistent with the view that late positive potentials reflect, at least in part, episodic memory encoding, although the subsequent memory effect provides a purer measure for this. Exploratory correlations also support a role for the LPP in memory formation (see Table 2). Topographic www.nature.com/scientificreports/ differences may indicate that the verbal learning group relied more on frontally mediated strategic mechanisms, particularly to support encoding of neutral items. However, given the modest sample sizes of each group, these relationships should be interpreted cautiously. Only in the slow-wave window did the verbal-learning group exhibit more positive-going ERPs than the other two groups. Given the role of positive-going slow waves in strategic verbal memory encoding 27 , this finding might explain why the verbal-learning group had similar recognition accuracy (and even more hits) than the levels-of-processing group, despite having smaller amplitudes on the earlier ERPs. Therefore, encoding in the social feedback and levels-of-processing groups seems to have relied more on automatic mechanisms, whereas the verbal-learning group recruited more strategic processes.
As expected, significant emotion effects on brain potentials were also observed, although they were smaller than the context effects. Emotion sensitivity of EPN and LPP components aligns with much previous research [e.g., see 17,[40][41][42] ]. Largest amplitudes for positive words have also been reported before [see 43 ]. Self-referential processing might have contributed to a more substantial increase for positive content, whereas in other situations, a bias towards negative material might occur [e.g., see 40,41 ].
Overall, we show that social context during encoding enhances long-term recognition memory beyond explicit learning or typical deeper incidental encoding tasks (self-descriptiveness or concreteness decisions). Notably, although both the social feedback and self-descriptiveness task probably recruited self-referential processing, effects in the feedback condition, where any self-reference was socially contextualized, by far exceeded those of task-driven self-referential processing alone. This finding may give rise to questions regarding the role of the initial structured interview and self-introduction, which were designed to make the entire social context more salient and credible. However, this manipulation also bears some similarities with common stress induction methods. We observed no specific increases in state anxiety ratings compared to the other groups. Trait anxiety ratings did not differ either. Still, the present implementation might be regarded as a mild form of stress induction and may have induced arousal or led to secretion of stress hormones which we did not measure. The existing literature on memory effects of pre-encoding stress suggests that the present effects are unlikely solely due to pre-encoding stress. Previously reported effects of pre-encoding stress 30 min before encoding (which would be quite similar to the present timing) on recognition memory the next day reveal modest 44 , ambiguous 45,46 , and often also impairing effects 47 which contrasts with the quite large effects that we found after a weeklong retention interval. Nevertheless, any endocrine effects of the present context induction should be addressed in future studies. Although the present retention interval of a week leaves ample time for consolidation or spontaneous rehearsal, which we cannot directly assess, the immediate repetition run revealed no significant effects, suggesting that effects were either due to initial encoding, where corresponding ERP modulations occurred, or arose considerably later.
The fact that responses in the social feedback and the levels-of-processing groups were self-paced may give rise to concerns regarding processing times. However, processing times were quite similar in these two experiments, and if anything, they were longer in the levels-of-processing group, yet memory performance was much higher in the social feedback group. The levels-of-processing group returned for testing on average half a day after the social feedback group, which might be another source of concern. Yet, there was no difference in testing lag between the verbal learning group and the social feedback group. The verbal learning group was the only group aware of the test, and they could use the interstimulus interval for rehearsal while the other two groups were performing the decision task. Still, the social feedback group clearly outperformed the other two groups regarding recognition memory.
Our focus was on the between-groups analysis, which de-emphasizes potentially interesting effects in individual conditions. Still, in Experiment 1, some "sender" differences occurred at encoding (see Supplement B), replicating previous research [e.g., see 28 ]. However, these had little effect on long-term recognition, suggesting that participants integrated both blocks from experiment 1 into one episode whose items were considerably more memorable than those from the other two experiments. Interactions in the levels-of-processing group indicate that encoding task modulates emotion effects in long-term recognition memory, only self-descriptiveness and not concreteness decisions resulting in an emotion effect on recognition accuracy, in line with the suggestion of interacting memory effects of self and emotion 16 .
In sum, although we used identical stimuli and presentation parameters across three experiments, the psychological encoding contexts elicited pronounced between-group differences in recognition memory and ERPs: The social-feedback context enhanced recognition memory beyond both explicit verbal-learning and self-descriptiveness/concreteness judgments. It also induced the largest EPN and LPP amplitudes, followed by levels-of-processing and verbal learning. The most positive-going slow wave amplitudes in the verbal learning group suggest that this group might have relied particularly on strategic memory processes. Our findings specify some important social factors in human long-term memory. They resonate with the fact that humans typically acquire information in social contexts [e.g., see 48 ], social evaluation being a particularly salient factor in human life [e.g., see 49 ]. Finally, these results have implications for memory formation in educational settings or memory assessment in legal contexts, where the extent to which individuals were exposed to an evaluative social context will affect their memory for elements of the initial episode.

Methods
Participants. Three groups of twenty-four participants each were recruited at Bielefeld University. Participants provided written informed consent and received 10 Euros per hour for participation. The Ethics Committee at Bielefeld University approved the study and the study was performed in accordance with the regulations of the Declaration of Helsinki. All subjects were right-handed, had normal or corrected-to-normal vision, and were free from a self-reported neurologic or psychiatric disorder. They were tested twice with a lag of about one www.nature.com/scientificreports/ week (T1-T2 difference M = 7.13 days, Min = 6 days, Max = 9 days, see Table 3). Experimental groups did not differ concerning BDI scores (F (2,68) = 0.03, p = 0.973, η p 2 = 0.001), or STAI state (F (2,68) = 1.34, p = 0.269, η p 2 = 0.038) and trait anxiety scores (F (2,64) = 0.72, p = 0.492, η p 2 = 0.022). There was a significant effect of the T1-T2 difference (F (2,64) = 7.30, p = 0.001, η p 2 = 0.186), showing a significant longer latency of about half a day for the levels-of-processing group compared to both the social-feedback group (p = 0.004) and verbal-learning group (p < 0.001). The social-feedback and verbal-learning groups did not differ (p = 0.464). Four participants in total did not attend the second session. One of these was in the levels-of-processing group and three in the verbal-learning group.
Stimuli. The stimulus set had been rated by 22 other students on nine-point Likert-type scales in terms of valence and arousal using the Self-Assessment Manikin 52 and a similarly constructed concreteness scale. The selected 270 adjectives (90 negative, 90 neutral, 90 positive) were assigned to three separate lists. Neutral adjectives were allowed to deviate from emotional adjectives on both arousal and concreteness since truly neutral trait adjectives are rare in an interpersonal evaluative context. The three lists did not differ in relevant emotional and lexical properties, and the list-condition assignment was counterbalanced (see Table 4).
Procedure. An outline of the experimental setup for the three groups is presented in Fig. 4.
Social-feedback group. Upon arrival, participants were told that another, unknown person, would evaluate them based on their self-presentation. In a second condition, a randomly operating computer algorithm was supposedly going to give them feedback. All subjects underwent both "sender" conditions. Condition sequence and word lists were counterbalanced. Importantly, random feedback was presented in both conditions. Participants were instructed to briefly describe themselves in a structured interview in front of a camera. They were Table 3. Demographic information for the participants in each experiment. (a) information is missing from one (b) missing from two (c) missing from three participants. *BDI: Beck's Depression Inventory [50]. **STAI: State-Trait Anxiety Inventory [51].  www.nature.com/scientificreports/ informed that the video of their self-description would be presented to another person. During EEG preparation, participants completed a demographic questionnaire and the BDI 50 and STAI 51 to characterize the sample. A research assistant left the testing room 15 min ahead of the fictitious feedback to ensure face validity, guiding an "unknown person" to a laboratory room next to the testing room. Stimuli were presented by software described as "Interactional Behavioral Systems", supposedly allowing instant online communication. Participants were told that the unknown other person would select adjectives describing the participant, which would be subsequently presented. In the other condition, feedback was denoted as random computer feedback. Feedback-adjectives were presented for 1500 ms, after which the participants had to indicate via button-press whether or not they accept this evaluation. After the response, a fixation cross was presented for 2000 to 3500 ms. For each sender, 30 negative, 30 neutral, and 30 positive adjectives (one of the three lists, see Table 4) were presented twice. The desktop environment and stimulus presentation were created using Presentation (www. neuro behav ioral syste ms. com). Afterwards, participants were debriefed that no social evaluation had taken place and presented a passive viewing run of all 60 negative, 60 neutral, and 60 positive adjectives again to test for any immediate post-processing effects of the manipulation.
Participants were informed that an unrelated test would be conducted in a second session one week later, where an unexpected recognition test took place. All adjectives from the first session and a set of new items were shown for 1500 ms each, followed by an old/new decision and a fixation cross for 1500 to 2000 ms.
Verbal-learning group. The very same stimuli and presentation parameters were used. In order to keep the visual input constant, the same background presentation setup (Interactive Behavioral Systems) was used but never referred to. The self-introduction phase was omitted, and no social significance was assigned to the stimulation. Instead, participants were instructed to memorize all adjectives for recognition testing in the next session one week later. Adjectives were presented for 1500 ms, followed by a variable fixation cross presented for 2000 Statistical analyses. For all data, we calculated three (group: social-feedback, verbal-learning, levels-ofprocessing) by two (condition: human sender/block A/self-reference, computer sender/block B /concreteness) by three (emotion: positive, negative, neutral) mixed ANOVAs. For memory data, the discrimination index (P r = hits-false alarms), and the response bias (B r = false alarms/(1 − P r )) were calculated according to Snodgrass and Corwin's two-high-threshold model 55 . Br values of 0.5 indicate no response bias, while higher values indicate a liberal and lower values a conservative response strategy. Raw data are presented in Supplementary Materials Section A. Furthermore, for the social-feedback and levels-of-processing groups, decision rates (agreement, selfdescriptiveness, concreteness) were assessed.
Finally, exploratory correlations between ERP amplitudes during encoding and memory performance were performed. To this end, mean ERP amplitudes were correlated with different memory indices (Hits and False Alarms (FA)) using JASP (www. jasp. org). We calculated both Bonferroni-corrected inferential and Bayesian Pearson correlation coefficients. For Bayesian analyses, the null hypothesis was specified as a point-null prior (i.e., standardized effect size δ = 0). It defined the alternative hypothesis as a Jeffreys-Zellner-Siow (JZS) prior, i.e., a folded Cauchy distribution centered around δ = 0 with the scaling factor r = 0.707. This scaling factor assumes a roughly normal distribution. To assign verbal labels to the strength of evidence, we followed the taxonomy suggested by Jeffreys 57 , labeling Bayes Factors with a BF 10 of 1 as no evidence, BF 10 between 1-3 as anecdotal evidence, 3-10 as moderate evidence, 10-30 as strong evidence, 30-100 as very strong evidence, and larger BFs as extreme evidence in favor of the alternative hypothesis.