Prediction Error-Related Memory Enhancement Depends on the Neural State Surrounding the Prediction Error Event

Abstract

Prediction errors (PEs) can enhance memory for preceding events. While such PE-related memory enhancements are critical for understanding adaptive memory, their underlying mechanisms are not fully understood. Using electroencephalography (EEG) and neuro-navigated transcranial magnetic stimulation (TMS) in combination with multivariate pattern analysis, this preregistered study aimed to elucidate the brain mechanisms underlying PE effects on memory. Specifically, we tested whether PEs trigger a neural reactivation of the preceding stimulus and whether the PE-induced effects on memory depend on the specific neural state before the PE. We also examined whether inhibitory TMS over the superior parietal cortex (SPC) reduces PE effects on memory. A total of 118 participants (male and female) received inhibitory theta-burst or sham stimulation over the SPC before completing an incidental encoding-fear learning task. In this task, participants learned trial-unique stimuli and predicted whether these would be followed by an electric shock, while EEG was recorded. Recognition memory was tested 24 h later. Our findings show that signed PEs enhance subsequent memory, depending on theta and alpha oscillations as well as neural category reactivation shortly before the PE. Moreover, this memory enhancement was associated with post-PE theta but not with PE-driven category reinstatement. Theta-burst stimulation over the SPC led to a more conservative mnemonic response bias but left the PE effect on memory unaffected. Together, our findings reveal that PE effects on memory formation are influenced by neural states and representations surrounding the PE, providing new insights into the neural mechanisms of adaptive memory formation.

Significance Statement

In daily life, our brain prioritizes events for memory storage that help predict relevant outcomes. Recent research revealed that prediction errors (PEs)—mismatches between anticipated and actual outcomes—linked to rewarding or aversive events enhance memory for preceding stimuli, highlighting their key role in adaptive memory formation. However, the underlying brain mechanisms remain unclear. This study demonstrates that the impact of PEs on memory depends on the neural category representation and brain oscillations shortly before and after the PE event. These findings provide novel insights into how neural states surrounding PEs influence memory formation, with potential implications for understanding maladaptive memory processes in fear-related disorders.

Introduction

Adaptive memory enables organisms to leverage past experiences to guide actions and choices (Shohamy and Adcock, 2010). However, not all events are stored equally well in memory, preference is rather given to information being crucial for predicting relevant outcomes. In support of this notion, research shows that prediction errors (PEs)—mismatches between expected and actual outcomes—associated with rewarding or aversive events enhance memory for preceding stimuli (Ergo et al., 2020; Kalbe and Schwabe, 2020; Rouhani and Niv, 2021; Rouhani et al., 2023). Although these PE effects are fundamental to our understanding of adaptive memory and may have significant implications for educational contexts and psychopathology, the brain mechanisms underlying the impact of PEs on memory for preceding events remain poorly understood.

Initial evidence from an fMRI study (Kalbe and Schwabe, 2022) suggests that PE effects on memory are associated with a reduced activation of the medial temporal lobe, implicated in memory formation for expectancy-congruent information (Davachi and Wagner, 2002; Eichenbaum, 2004), and an increased crosstalk of the salience and a frontoparietal network. While fMRI provides high spatial resolution to identify relevant brain areas, its temporal resolution is limited, making it difficult to capture neural processes occurring around the PE. Specifically, two short-lived neural mechanisms might drive PE effects on memory. First, PEs could evoke a transient reactivation of the preceding predictive stimulus, promoting its memory storage. This mechanism aligns with evidence indicating that postencoding reactivation is essential for subsequent recall (Staresina et al., 2013; Tambini and D’Esposito, 2020). A second mechanism may involve the neural state just before the PE. PEs might strengthen memory for preceding events, if these events are still neurally maintained when the PE occurs. This is in line with synaptic or behavioral tagging models proposing that preactivated representations can be enhanced by a subsequent salient event, such as a PE (Moncada et al., 2015). Potential candidate mechanisms for neural maintenance include alpha oscillations, implicated in attention to task-relevant stimuli (Payne and Sekuler, 2014), theta oscillations, related to the reinstatement of memory representations or the binding of associative memory (Nyhus and Curran, 2010; Staudigl and Hanslmayr, 2013; Kota et al., 2020), and the neural reactivation of the representation of the preceding stimulus.

If PE effects on memory require the reactivation or maintenance of the preceding stimulus around the PE event, this raises the question of whether interference with these mechanisms could reduce or even abolish PE effects on memory. The superior parietal cortex (SPC) has been repeatedly shown to be crucial for stimulus maintenance, working memory processes, and top-down attentional updating (Corbetta et al., 1995; Wager and Smith, 2003; Koenigs et al., 2009; D’Esposito and Postle, 2015; Ester et al., 2015), making it a promising candidate for the maintenance of the predictive stimulus. Thus, we hypothesized that inhibiting SPC functioning would reduce PE effects on memory for preceding stimuli.

In this preregistered study, we combined “neuro-navigated” transcranial magnetic stimulation (TMS) with electroencephalography (EEG) and multivariate pattern analysis to elucidate the brain mechanisms underlying PE effects on memory. Specifically, we tested whether (1) PEs induce a neural category reactivation of the preceding stimulus; (2) PE effects on memory require a specific neural state, associated with alpha or theta oscillations or a neural representation of the preceding stimulus shortly before the PE; and (3) inhibitory stimulation over the SPC reduces PE effects on memory. To these ends, we applied continuous theta-burst stimulation (cTBS) over the SPC before participants completed a combined incidental encoding-fear learning task, while EEG was recorded. During this task, participants encoded trial-unique stimuli and predicted whether these would be followed by an electric shock. Memory was tested 24 h later. We hypothesized that (signed) PEs would enhance subsequent memory and that these effects would be dependent on the neural state and representation around the PE. Additionally, we predicted that cTBS over the SPC would generally reduce PE effects on memory.

Materials and Methods

Preregistration

This study was preregistered at the German Clinical Trials Register (DRKS-ID: DRKS00030529; https://drks.de/search/en/trial/DRKS00030529).

Participants

One hundred twenty-two healthy right-handed volunteers participated in this study (69 female; age: M = 25.55 years, SD = 3.63 years, range = 19–33 years). Exclusion criteria were screened in a standardized interview and comprised the following: insufficient command of German, lifetime history of any neurological, cardiovascular or psychiatric diseases, medication intake or substance abuse, and contraindications for MRI measurements or TMS. All participants provided written informed consent before participation and received a monetary reimbursement. The ethics committee of the Faculty of Psychology and Human Movement Science at the University of Hamburg approved the study (2022_055_Loock_Schwabe), which was carried out in line with the Declaration of Helsinki.

The target sample size was based on a previous behavioral study from our lab that showed an effect of aversive signed PE on memory formation using the same task in n = 120 participants with a power of 0.92 (Loock et al., 2025). In the present study, a post hoc power simulation using the R-package simR (Green and MacLeod, 2016) for the observed effect of aversive signed PE on memory formation and our final sample size of n = 118 participants (due to exclusions in EEG data analysis) yielded a power of 0.98 based on 1,000 simulations.

We employed a between-subjects design with the factor stimulation group (TMS vs sham). Participants were randomly assigned to the TMS group (n = 62, 31 female) and the sham group (n = 60, 38 female).

Experimental design

The experiment consisted of one MRI session—during which we acquired an anatomical brain image required for “neuro-navigated” TMS—and two experimental sessions (Day 1 and Day 2) that took place on 2 consecutive days (Fig. 1).

Figure 1.

Overview of the experimental procedure. In the delayed-matching-to-sample task, participants were required to keep stimuli from three different categories (animals, scenes, tools) in mind for 2 s and to select which stimulus they had seen before, which was then used for classifier training. TMS/Sham stimulation was applied to the right superior parietal cortex (highlighted in red). In the incidental encoding-fear learning task, participants saw a series of trial-unique pictures from three different categories (animals, scenes, tools) linked to fixed probabilities for receiving an aversive electric shock (CS^a+ = 67%, CS^b+ = 33%, and CS⁻ = 0%). On each trial, participants indicated their shock expectation on a continuous scale (from 0 to 100%). The delay with which the outcome (shock vs no-shock) occurred after stimulus offset varied between 0 and 10 s. In a surprise recognition test 24 h later, participants had to indicate whether they had seen the item on the screen before while indicating their certainty (definitely old, maybe old, maybe new or definitely new). Importantly, they were presented with old items from the incidental encoding-fear learning task and unseen, new items. Critically, the TMS/Sham stimulation over the right superior parietal cortex was applied before and during the incidental encoding-fear learning task. All depicted images are licensed under Creative Commons BY-SA license.

Before experimental Day 1, we acquired T1-weighted structural magnetic resonance (MR) images of each participant using a 3T Siemens PRISMA scanner located at the University Medical Center Hamburg-Eppendorf. We utilized a magnetization-prepared rapid acquisition gradient echo (MPRAGE) sequence to collect the anatomical images that had a voxel size of 0.8× 0.8 × 0.9 mm³ and consisted of 256 slices. The imaging parameters for the MPRAGE sequence were a repetition time (TR) of 2.5 s and an echo time (TE) of 2.12 ms. These structural brain images were used for “neuro-navigating” the TMS or sham stimulation.

Upon arrival on experimental Day 1, participants provided written informed consent and filled out questionnaires assessing depressive symptoms (BDI-II; Beck et al., 1961), sleep quality (PSQI; Buysse et al., 1989), state and trait anxiety (STAI-S and STAI-T; Spielberger et al., 1970), and chronic stress (TICS; Schulz et al., 2004). While completing the questionnaires, the EEG cap and electrodes were set up. A stimulation electrode for applying electric shocks during the incidental encoding-fear learning task was placed on the participant's right lower leg, ∼20 cm above the heel. Shock intensity was adjusted individually to be unpleasant but not painful in a stepwise manner. More specifically, 200 ms single pulse shocks were administered consecutively with an initial intensity of 15 V until they were perceived as unpleasant but not painful. For electrical stimulation, we used the STM-200 stimulation module connected to the MP-160 data acquisition and analysis system (BIOPAC Systems). To assess skin conductance responses (SCRs) as an indicator of physiological arousal, electrodes were placed on the distal phalanx of the index finger and the third finger of the left hand. Next, participants’ individual motor-thresholds for TMS were determined. Thereafter, participants performed the first session of a Delayed-matching-to-sample (DMS) task (see below), which served to later train a classifier based on L2-penalized logistic regression for EEG-based decoding, before they underwent either the sham or TMS stimulation targeting the right SPC. Immediately after the TMS or sham stimulation, participants performed a combined incidental encoding-fear learning task in which they were asked to predict whether a stimulus presented on screen would be followed by an electric shock (see below). After half of the task, we administered a second TMS or sham stimulation to maintain the effects of the stimulation throughout the task. After finishing the encoding task, participants performed a second session of the DMS task. In total, Day 1 took ∼4.5 h per participant. Approximately twenty-four hours later, participants returned for a surprise recognition test in which we assessed their memory for the stimuli encoded on Day 1.

Day 1: delayed-matching-to-sample task

In order to decode neural stimulus representations before and after PE, we trained a classifier based on the EEG data from a DMS task (Meier et al., 2022; Fig. 1). The DMS is common for examining working memory processes (Anderson and Colombo, 2019). This task was performed twice, once before TMS/sham stimulation and once after the combined incidental encoding-fear learning task to rule out any time- or TMS-related biases in the classifier. In each of the two sessions, participants completed 150 trials. During each session, participants saw images of three different stimulus categories (animals, scenes, tools). Stimuli were taken from available image databases, i.e., Bank of Standardized Stimuli (Brodeur et al., 2010, 2014), SUN database (Xiao et al., 2010), Konklab (Konkle et al., 2010), and open online sources. In total, the stimulus set consisted of 300 unique pictures of animals, scenes, and tools, respectively, with 100 images per category isolated on white background. We preallocated two different stimulus sets of 150 pictures each (50 animals, 50 scenes, 50 tools) of which one was used in the first DMS session and the other in the second DMS session. All stimuli were assumed to be emotionally neutral and represented a unique exemplar of its category. On each trial, a target stimulus was presented for 2 s in the center of a gray screen. Participants were instructed to keep this trial-specific target in mind for a 2 s delay period during which a black fixation cross was presented on the screen. After the delay period, the target stimulus and two distractor stimuli appeared on the screen simultaneously, and participants were required to select via button press within 2 s which stimulus they had seen before. The position of the target stimulus on the screen (left, center, right) was randomized across trials and response buttons corresponded to the numbers “1” (left), “2” (center), and “3” (right) on the keyboard. The distractors were either drawn randomly from the same category as the target or from the two leftover stimulus categories. Trials were pseudorandomized with the restriction that successive trials did not include targets from the same category for more than three consecutive times. Between trials, there was a fixed interval of 2 s. An example trial of the DMS task is presented in Figure 1. Importantly, the stimuli used in the DMS task did not overlap with those used in the incidental encoding-fear learning task.

Day 1: TMS and sham stimulation

In order to examine the functional role of the SPC in PE-induced memory enhancements, we used neuro-navigated TMS over the right SPC before participants underwent the incidental encoding-fear learning task. For stimulation, we used a PowerMag Research 100 stimulator (MAG & More) which applies repetitive transcranial magnetic stimulation (rTMS). Depending on the experimental group (TMS vs sham), two different figure-eight TMS coils were used: a PMD70-pCool coil (MAG & More; maximum magnetic field strength of 2T) was used for continuous theta burst stimulation (cTBS) in the TMS condition, whereas the PMD70-pCool-SHAM (MAG & More; minimal magnetic field strength) was used in the sham condition. Importantly, the sham condition induced a similar sensory experience on the scalp not pervading the skull. We used a double-blind protocol, in which neither the participant nor the experimenter was aware of the stimulation condition. Participants were asked to guess which treatment they had received at the end of the experiment.

Motor threshold determination

The motor threshold (MT) was determined before participants started to do any task on Day 1 but were already wearing an EEG cap without electrodes attached to it (Grob et al., 2024). The MT determination was used to determine the appropriate magnetic field strength per individual. Disposable, pre-gelled Ag/ACL surface electromyography (EMG) electrodes were attached to the participant's right hand: An active electrode was placed on the abductor pollicis brevis muscle, with a reference electrode on the bony landmark of the index finger and a ground electrode on the tip of the ulna bone. In order to locate the motor hotspot (MH), we located the center of the head, moved 5 cm leftward and 3.5 cm to the forehead at an angle of 45° and marked this area as the center of a 3 × 3 point-grid area. Each point was 1 cm apart from its neighbors. Starting at 40% maximum stimulator output, we gradually increased the output intensity of the stimulation (with a step size of 5%) while adjusting the TMS coil to an angle of 45° on the z-axis. Then, we screened the 3 × 3 search grid for the motor hotspot delivering single 10 Hz pulses. As soon as the MH was found, the MT was determined at that certain location. The MT was defined as the minimum percentage of maximum stimulator output over the left motor cortex (area: M1) necessary to elicit motor evoked potentials (MEPs) with a peak-to-peak amplitude of 50 µV in response to at least eight out of 16 consecutive single pulses.

Neuro-navigation

Individual T1-weighted anatomical MR images of the participants were used for neuro-navigation with the PowerMag System (MAG & More). This procedure ensured a precise and individually tailored coil placement being aligned with the SPC as target area. An infrared camera (Polaris Spectra) was used to locate and track the participant's head and the TMS or sham coil in space. Based on the information we gained from the T1-weighted MR images, we created 3D models of the participant's head which allowed us to precisely locate the right SPC individually based on Talairach (TAL) coordinates from previous work (TAL: 21, −54, 51; Ester et al., 2015) which we transformed into system-compatible MNI coordinates (20, −58, 57). As we assume that memory maintenance processes might be involved in PE-related memory enhancements, we decided to target the SPC which has been repeatedly associated with working memory and stimulus representation (Koenigs et al., 2009; D’Esposito and Postle, 2015; Ester et al., 2015). After entering the target MNI coordinates (20, −58, 57), the coil was positioned in alignment with the neuro-navigation system. We aimed for a brain-to-target distance of <3 cm to ensure the shortest distance to the cortex.

Stimulation protocol

We applied cTBS using the active coil for the TMS group or the sham coil for the sham group. It is assumed that cTBS leads to an inhibitory effect on the target brain region under stimulation (Huang et al., 2005; Jannati et al., 2023; Grob et al., 2024). Based on a cTBS protocol by Grob et al. (2024), participants received a series of theta bursts with three magnetic pulses (triplets) at a frequency of 50 Hz, with the triplets being repeated at a rate of 5 Hz, i.e., five triplets per second. In total, 600 magnetic pulses over 40 s per participants were administered to the target area. We fixed the coil using a tripod placed behind the participant to maintain a precise (TMS or sham) stimulation of the right SPC (MNI: 20, −58, 57) with <3 cm of brain-to-target distance.

Day 1: incidental encoding-fear learning task

To examine the neural mechanisms underlying PE-induced memory enhancements, participants completed an incidental encoding-fear learning paradigm immediately after TMS or sham stimulation while EEG was recorded (Fig. 1). Stimuli were taken from existing databases, i.e., Bank of Standardized Stimuli (Brodeur et al., 2010, 2014), SUN database (Xiao et al., 2010), Konklab (Konkle et al., 2010), and open online sources. The stimulus set consisted of 540 pictures of animals, scenes, and tools, with 180 pictures per category isolated on white background. All stimuli were assumed to be of neutral valence and represented a unique exemplar of its category. Out of this pool, 360 pictures (120 per category) were randomly drawn and used during encoding on Day 1. The remaining 180 pictures (60 pictures per stimulus category) served as foils in the recognition test on Day 2. The order of item presentation was randomized across participants.

In the incidental encoding-fear learning task, participants were instructed that they would see a stream of pictures (animals, scenes, tools) presented one after another on the screen and that some pictures will be followed by an electric shock. Participants were asked to predict how likely a shock would be to follow the presented picture by adjusting a slider on the screen to a value that corresponded with their prediction of the shock probability (range, 0–100%). For each trial, we derived a PE which was calculated as the relative value of the difference between participants’ continuous explicit shock expectancy ratings (ranging from 0, corresponding to full confidence that no shock would occur, to 1, corresponding to full confidence that a shock would occur) and the actual binary outcome of the trial (coded “0” if no shock occurred and coded “1” if a shock occurred in the current trial). Importantly, participants were neither informed about the true shock contingencies, nor about the recognition test on Day 2. They were informed that the shock occurrences were not affected by their predictions but that they should learn by trial-and-error to improve their predictions over the duration of the task. Unbeknownst to the participants, the probabilities of a shock were linked to the three picture categories. One category served as CS^a+ (67% shock probability), one as CS^b+ (33% shock probability), and one as CS⁻ (0% shock probability). The assignment of image categories (i.e., animals, scenes, tools) to the CS categories (i.e., CS^a+, CS^b+, CS⁻) was counterbalanced across participants and groups, ensuring that our results are not confounded by stimulus category differences. Throughout the incidental encoding-fear learning task, SCR was measured as an indicator of physiological arousal by using electrodes on the individual's left hand. Electric shocks were applied via the shock electrode placed at the participant's lower right leg.

In total, the incidental encoding-fear learning task consisted of 360 trials split into four blocks of 90 trials. In each block, 30 pictures of animals, 30 pictures of scenes, and 30 pictures of tools were presented in a pseudorandomized order, so that no more than three pictures of the same category appeared in a row. On each trial, a picture was shown in the center of the screen for 4.5 s, during which participants were asked to make their prediction about the probability of an electric shock (Fig. 1). A slider was presented underneath each item which could be individually adjusted to any integer value between 0 and 100% by using the computer mouse. After stimulus offset, a black dot appeared centrally on the screen which coterminated with the 200 ms outcome (shock vs no-shock), i.e., in no-shock trials, the transition from dot to fixation cross indicated that the CS was not followed by a shock. Critically, the duration of the dot's presentation on the screen ranged randomly between 0 and 10 s per trial to vary the critical CS-outcome delay. After the outcome, a black fixation cross centered on gray background was presented for 6.5 ± 1.5 s. Between blocks, there was a short break (1–2 min) during which participants had the chance to recalibrate the shock intensity and rest, if required. Each encoding block lasted ∼25 min, resulting in a total duration of 100 min for the entire incidental encoding-fear learning task.

Our experimental design allowed us to capture continuous predictions which resulted in continuous PEs. Based on recent literature on PEs (Kalbe and Schwabe, 2022; Rouhani et al., 2023; Loock et al., 2025), we calculated signed PEs (sPE) which facilitated the distinction between positive and negative PEs. The sPE was calculated as the relative difference between the binary outcome (shock vs no-shock) and the explicit shock prediction in the respective trial resulting in a value between −1 and 1. For example, if a shock occurred, i.e., outcome = 1, but the predicted shock probability was 0.7 (corresponding with 70%), the resulting sPE would be 0.3 (1–0.7 = 0.3). Conversely, if the shock was omitted, i.e., outcome = 0, the sPE would be −0.7 (0–0.7 = −0.7). Notably, the sPE's sign also indicated the value of the outcome: Negative sPEs (sPE < 0) could only occur in unshocked trials, i.e., unexpected shock omissions, while positive sPEs (sPE > 0) could only occur in shocked trials, i.e., unexpected shock occurrence. Additionally, we derived unsigned PEs (uPEs) which were calculated as the absolute value of the difference between participants’ continuous explicit shock expectancy ratings and the actual binary outcome of the trial (shock vs no-shock). The resulting unsigned PE is, therefore, ranging between 0 and 1.

Day 2: recognition memory test

On experimental Day 2, ∼22–26 h after Day 1, participants returned for a surprise recognition test (Fig. 1). First, they completed a short questionnaire to assess whether they anticipated a memory test and then rated how surprised they were about the recognition test on a scale from 1 (“not surprised at al”) to 5 (“very surprised”). In the recognition test, participants saw all pictures they had seen during the incidental encoding-fear learning task (120 pictures of animals, 120 pictures of scenes, and 120 pictures of tools) as well as 180 “new” pictures, i.e., foils (60 pictures of animals, 60 pictures of scenes, and 60 pictures of tools) that had not been presented before, in a randomized order. In line with prior memory studies (Staresina and Davachi, 2006; Gimbel and Brewer, 2010; Kim et al., 2014), we used half as many foils as targets (2:1 target ratio) to reduce overall task duration and to mitigate potential fatigue effects, particularly given the length and cognitive demands of the experimental procedure.

Each trial started with a centered black fixation cross on a gray background for 1.5 ± 0.5 s, followed by an “old” or “new” picture presented centrally on the screen for 6 s. For each item, participants were instructed to indicate whether the currently presented picture was definitely old, maybe old, maybe new, or definitely new by pressing the “1,” “2,” “3,” or “4” button on the keyboard, respectively. Participants had to log in their response while the stimulus was presented on screen (maximum 6 s).

SCR data acquisition and analysis

On experimental Day 1, we recorded SCR as a measure of arousal and conditioned fear during the incidental encoding-fear learning task. SCR was measured using a MP-160 BIOPAC data acquisition system (BIOPAC Systems).

SCRs were analyzed using Continuous Decomposition Analysis in Ledalab version 3.4.9 (Benedek and Kaernbach, 2010). On each trial, we derived the average phasic driver within a specified response window. First, the skin conductance signal was downsampled to 50 Hz and optimized applying four cycles of initial values to increase the goodness of the model. For the anticipatory SCR, a response window was set from 0.5 after stimulus onset until the onset of the outcome (shock/no-shock) and could vary depending on the CS-outcome delay. Outcome-related SCR was analyzed between 0.5 and 4.5 s after outcome onset. The minimum amplitude threshold was set to 0.01 μS for both the anticipatory and the outcome-related SCR. Resulting estimates of the average phasic driver within each response window were returned in microsecond. Notably, these estimates are sensitive to interindividual differences because of physiological factors such as the thickness of the corneum (Figner and Murphy, 2011). We therefore standardized both the anticipatory and the outcome-related SCR by dividing the average phasic driver estimated in each trial by the maximum average phasic driver for each participant observed in every trial. During SCR analysis, we noticed that there were no significant SCRs to either the unconditioned stimulus, i.e., an electric shock, or the CS. This was likely due to combination of a high proportion of participants with weak or absent electrodermal responsiveness and technical failure, which rendered the SCR data unreliable. We therefore decided to not include the SCR data in further analyses.

EEG data acquisition and analysis

EEG acquisition

On Day 1, EEG was recorded during each of the DMS sessions and the incidental encoding-fear learning task. Participants were seated 80 cm in front of a computer screen in an electrically shielded and sound-isolated room. A 64-channel BioSemi ActiveTwo system (BioSemi), following the international 10–20 system, was used to record EEG at a sampling rate of 1,024 Hz. Additional electrodes were placed at the mastoids, above and below the orbital ridge of the right eye and at the outer canthi of both eyes. Electrode impedances were kept between ±30 mV. The EEG data was referenced online to the BioSemi common mode sense (CMS)-driven right leg (DRL) reference electrodes and filtered online with a bandpass filter of 0.03–100 Hz. Due to technical issues, four participants had to be excluded from EEG analysis resulting in a sample of n = 118 participants for the EEG analysis.

Preprocessing

Preprocessing was performed offline using the FieldTrip toolbox (version 20200607; Oostenfeld et al., 2011) and custom scripts in Matlab (version 2020b; The MathWorks). Trials from the incidental encoding fear-learning task were segmented from −5 to 5 s relative to outcome onset and rereferenced to the mean average of all scalp electrodes. Data were demeaned based on the average signal of the entire trial and detrended. A discrete Fourier-transform filter (DFT) at 50 Hz was applied to minimize power-line noise. Electrodes that did not record or showed extensive noise (maximum of one per participant) were removed and interpolated by weighted neighboring electrodes. Noisy trials were removed by visual inspection. On average, 11.05 (SD = 5.57) of the 360 trials were removed in the incidental encoding-fear learning task, corresponding to ∼3% of all trials. After artifact rejection, the segments were downsampled to 256 Hz. Next, we ran an extended infomax independent component analysis (ICA; Makeig et al., 1995) using the “runica” method with a stop criterion of weight change <10^–7 in order to identify and reject components associated with eyeblinks or other sources of noise (Oostenveld et al., 2011). Following a two-step procedure, we first correlated the signals from the horizontal and vertical EOG electrodes with each independent component and removed components exhibiting a correlation higher than 0.9. The remaining components were then identified through visual inspection along the time course and corresponding brain topographies. On average, 6.08 (SD = 2.67) components per participant were detected and removed.

Event-related potential analysis

Based on previous studies on PE-related event-related potentials (ERPs; Silvetti et al., 2014; Turan et al., 2025), we analyzed outcome (i.e., shock vs no-shock) effects on the feedback-related negativity component (FRN; Talmi et al., 2013) and on the Positivity 300 component (P3; Ridderinkhof et al., 2009) in the EEG data of the incidental encoding-fear learning task on Day 1. For this ERP analysis, data was segmented into epochs from −2,000 to 2,000 ms relative to outcome onset (shock vs no-shock) and baseline-corrected by subtracting the average 2,000 ms interval before outcome onset. The FRN was analyzed between 0 and 800 ms after outcome onset at the frontocentral electrode channel FCz independent of the trial outcome (shock vs no-shock) and defined as the largest negative peak between 0 and 800 ms after outcome onset used for later analyses with item recognition and PE effects (Fig. 2). The P3 was analyzed between 300 and 1,000 ms at the posterior channel Pz independent of the trial outcome and defined as the largest positive peak between 300 and 1,000 ms after outcome onset used for later analyses with item recognition and PE effects (Fig. 2).

Figure 2.

Overview of the analytical approach. We used a combination of linear mixed models (LMMs) and generalized linear mixed models (GLMMs) to investigate how trial-specific electrophysiological measures relate to subsequent memory and PE effects. During the pre-outcome window, i.e., 2 s before the outcome, we analyzed theta and alpha oscillatory power, category reactivation of the conditioned stimulus (CS), and their interaction with PEs, as induced by the outcome (shock vs no-shock) in relation to subsequent item recognition. In a 2 s time window after the outcome, we performed analogous analyses for outcome-evoked theta and alpha power, CS category reactivation, and event-related potentials (ERPs), i.e., the FRN and P3, and their respective contributions to subsequent memory. The depicted image is licensed under Creative Commons BY-SA license.

Time–frequency analyses

EEG data from the incidental encoding-fear learning task was decomposed spectrally using sliding Hanning windows (2–30 Hz, 1 Hz steps, five-cycle window, interval: −5 to 5 s relative to outcome onset) averaged over all trials. This enabled us to calculate the time–frequency representations with respect to two temporal windows: pre-outcome (−3 to 0 s relative to outcome onset) and outcome-evoked (0–3 s relative to outcome onset). To obtain a more nuanced picture of what is emerging around the occurrence of a PE, we also computed shorter time windows (pre-outcome: −1 to 0 s relative to outcome onset; outcome-evoked: 0–1 s relative to outcome onset) for which we obtained a similar pattern of results. In each window, trial-wise power estimates were calculated, log-transformed to reduce skewness and increase normality of the data (Grandchamp and Delorme, 2011; Smulders et al., 2018), and baseline corrected (absolute baseline correction −5 to −3 s relative to stimulus onset). For the whole-brain time–frequency data, spectral power averaged over all trials was tested with a dependent sample cluster-based permutation t test (10,000 permutations to correct for multiple comparisons; Maris and Oostenveld, 2007). This approach allows testing for statistical differences while simultaneously controlling for multiple comparisons without spatial constraints. The samples were clustered at a level of α_cluster = 0.001. Clusters with a corrected Monte Carlo p value < 0.05 are reported as significant. Additionally, we entered the single power estimates in alpha (8–12 Hz) and theta bands (4–7 Hz) into models assessing item recognition dependent on PEs (Fig. 2).

MVPA

Multivariate decoding analysis was performed using the MVPA-light toolbox (Treder, 2020).

Classifier training

For decoding, EEG data from the DMS sessions before and after the incidental encoding-fear learning task was pooled to reduce any time-related biases and to make sure that there is a sufficient number of trials for a reliable classifier training. In the DMS task, epochs were defined as −2,000 to 2,000 ms relative to the delay phase. Then, the EEG data was processed exactly as in the PE-task. On average, 7.76 (SD = 3.42) of the 300 trials (pooled over both sessions) were removed from the DMS task, corresponding to <3% of all trials. During ICA, 4.97 (SD = 2.20) components per participant were detected and removed on average. Afterward, the classifier was trained within-subject, utilizing a logistic regression (L2 penalized) on the preprocessed data of the DMS task (pooled over both DMS sessions) to differentiate between image categories (animals, scenes, tools). On each trial, each target category (e.g., a scene) was contrasted against the two unseen image categories (e.g., an animal and a tool). To account for class imbalances, we applied class weights that incorporated the inverse frequency of each class.

All EEG channels were used as features. Prior to classification, we segmented the preprocessed data into one time window that was subject of the classifier training: The investigated window (0–2,000 ms relative to the delay phase) included the whole delay phase where participants had to keep the target in mind (stimulus maintenance phase). Here, we used a sliding window averaging 100 ms with a step size of 10 ms to identify the optimal time window for individual decoding. To evaluate the classification performance, we implemented a fivefold cross-validation. The classifiers with the highest performance, i.e., accuracy, were used to decode the neural representations during the incidental encoding-fear learning task.

Decoding

The classifier trained during the stimulus maintenance phase per participant was used to decode neural patterns emerging before and after outcome onset in the incidental encoding-fear learning task, i.e., before and after the occurrence of PEs. As the classifier was trained to distinguish between stimulus categories, decoding results are limited to category-level information and do not permit conclusions about stimulus-specific representations. We defined two windows of decoding and segmented the preprocessed data from the incidental encoding-fear learning task accordingly: A pre-outcome window during the CS-outcome delay (−2,000 to 0 ms relative to outcome onset) and an outcome-evoked window after the outcome was presented (0–2,000 ms relative to outcome onset). The pattern of results remained unchanged when analyzing smaller time windows (±1,000 ms relative to outcome onset). The pre-outcome window was indicative of the neural stimulus representation (i.e., maintenance) during the CS-outcome delay, shortly before a PE occurred, while the outcome-evoked window indicated neural patterns that emerge after a PE has occurred (i.e., potential category reactivation). In a trial-wise manner, the classifier trained during stimulus maintenance in the DMS task was applied to each of the decoding windows (pre-outcome, outcome-evoked) utilizing an overlapping sliding window, with a time average of 100 ms and a step size of 10 ms. The resulting average decoding accuracy per trial indicated the strength of the neural patterns before and after the a PE, respectively, and was used for later analyses of item recognition and PE effects (Fig. 2).

Statistical analysis

Behavioral analyses

Overall, item recognition was treated as the binary dependent variable, coded as “0” for misses and “1” for hits. In line with previous research on episodic memory (Bartlett et al., 1980; Gagnon et al., 2019; Kalbe and Schwabe, 2022; Heinbockel et al., 2024), our analysis focused on high-confidence responses, such that only trials in which participants indicated that they were “very sure” were considered as hits. Such high-confidence recognitions have been linked to a hippocampus-based recollection rather than only familiarity with an item, which is assumed to depend on the perirhinal cortex (Eichenbaum et al., 2007). Accordingly, we computed hit rates (i.e., recognizing an item as “surely old”) and category-based false alarm rates based on stimulus category-level (CS^a+ vs CS^b+ vs CS⁻). Additionally, we computed the signal detection theory-based parameter d′, computed as the difference between z-transformed hit rates and z-transformed false alarm rates, where z represents the inverse of the standard normal distribution. Using mixed-design ANOVAs, we also examined the influence of the between-subjects factor stimulation group (TMS vs sham) and the within-subject factor conditioning category (CS^a+ vs CS^b+ vs CS⁻).

Linear and multilevel models

To analyze how PEs impacted subsequent recognition memory, we fitted generalized linear mixed models (GLMMs) with a logit link function using the lme4 R package (Bates et al., 2015) that enabled us to perform trial-wise analyses. To maximize generalizability of the GLMMs, we utilized the maximal random effects structure and treated subjects as random effects for both the intercept and all slopes of the fixed effects included in the model (Barr et al., 2013). The recognition of an individual item was treated as the binary dependent variable, coded “0” for misses and “1” for confident hits. For PEs, we derived a fine-grained measure of PEs, the sPE, ranging between −1 and 1, and allowing to differentiate between negative and positive PEs, while being treated as a continuous variable in all statistical models.

We fitted models using different sets of independent variables, including sPEs, the explicit shock prediction, the CS-outcome delay, and the stimulation group including their interaction. The best fitting models were selected based on χ² tests. Our analytical approach followed a stepwise, hypothesis-informed process (Anderson and Burnham, 2004; Zuur et al., 2009) in which we added regressors based on our a priori research questions to identify the most parsimonious models. Model comparisons (between sPEs and uPEs) and formulas for the main analysis models are provided in Table S1 in the supplemental material.

To directly link electrophysiological data with the participant's recognition performance and PE effects, we computed a set of separate linear mixed models (LMMs) and GLMMs at the trial level. For ERP data, we built LMMs that predicted either FRN amplitude or P3 amplitude from the independent variables sPE, the stimulation group, CS-outcome delay including the sPE × CS-outcome delay, sPE × stimulation group, and sPE × CS-outcome delay × stimulation group interaction, and the explicit shock prediction. Additionally, we also computed GLMMs that tested whether the P3 amplitude and the FRN amplitude, respectively, predicted item recognition (Fig. 2).

For time–frequency data, we fitted LMMs to investigate whether the average spectral power interacted with the PE in the pre-outcome and outcome-evoked window to affect item recognition. For each window, we computed a set of GLMMs that included the average spectral power, sPE, CS-outcome delay, and stimulation group including their interactions to predict item recognition (Fig. 2). Notably, we investigated pre-outcome and outcome-evoked windows that lasted 3 s.

To analyze if and how the neural stimulus representations are associated with PE-induced memory enhancements, we also computed (G)LMMs in which we included decoding accuracy from the MVPA. To examine to what extent PE effects on subsequent memory require a neural stimulus (category) representation at the time of the outcome (i.e., shortly before a PE occurred), we computed a GLMM with item recognition being predicted by sPE and decoding accuracy (and stimulation group) in the pre-outcome window (Fig. 2). In order to assess whether PE magnitude is associated with category reactivation, we set up an LMM in which we predicted the outcome-evoked decoding accuracy by the sPE and the stimulation group. Additionally, we fitted a GLMM with the outcome-evoked decoding accuracy, sPE and stimulation group as independent variables to predict the binary variable item recognition and to further elucidate the mechanisms underlying PE-induced memory enhancement (Fig. 2).

All analyses were performed in R Studio [version 1.2.5033, RStudio Team (2020), PBC], unless indicated otherwise above, and subjected at a significance level of α = 0.05 and reported p values are two-tailed. In case of sphericity violation, indicated by Mauchly's test, Greenhouse–Geisser corrected degrees of freedom and p values are reported. Post hoc tests following significant main or interaction effects were Bonferroni corrected for multiple comparisons, if required. Effect sizes were either reported as Cohen's d for t values for between-subjects analyses, Cohen's d_z for within-subjects analyses, or partial eta squared for F values.

Data and code accessibility

All data, materials, and scripts have been made publicly available and can be accessed at https://doi.org/10.25592/uhhfdm.17016. This study was preregistered at the German Clinical Trials Register (DRKS-ID: DRKS00030529; https://drks.de/search/en/trial/DRKS00030529).

Results

Successful fear learning

PEs were overall equally distributed around zero indicating that a sufficient number of positive and negative PEs could be analyzed (Fig. 3A). Participants’ explicit shock predictions (ranging from 0 to 100%) showed that they learned the shock contingencies very well (Fig. 3B). For CS^a+ (M = 0.68, SD = 0.16), the shock expectancy was significantly higher compared with CS^b+ (M = 0.47, SD = 0.15; t₍₁₂₁₎ = 9.93, p < 0.001, d = 1.36) and for CS^b+ compared with CS⁻ (M = 0.06, SD = 0.13; t₍₁₂₁₎ = 23.33, p < 0.001, d = 3.22). Importantly, shock predictions did not differ between stimulation groups (F_(1,120) = 0.06, p = 0.806, partial η² = 0.00).

Figure 3.

PEs, shock expectations, and memory performance. A, PEs for CS^a+ and CS^b+ were equally distributed around zero. B, Participants' mean shock predictions (thick lines) approached the underlying shock probabilities (dotted lines) relatively quickly confirming successful fear learning. C, D, Memory sensitivity, computed as d-prime (computed as the difference between z-transformed hit rates and z-transformed false alarm rates) in the sham group (C) was significantly lower than in the TMS group (D). Dots show data from individual participants. E, sPEs significantly boosted memory for the preceding stimulus.

General memory performance

Recognition memory performance was assessed using the signal detection theory-based parameter d’ indexing memory sensitivity. Participants demonstrated robust memory performance across conditions, with higher d′ values for CS^a+ and CS^b+ items compared with CS^– items. Additionally, participants who received cTBS over the SPC showed enhanced memory sensitivity (d′) relative to those in the sham group. In a follow-up, we analyzed hit and false alarm rates separately revealing condition-specific response patterns and differences in response biases between groups.

When using the signal detection theory-based parameter d′, recognition memory was higher for both CS^a+ items (M = 0.90, SD = 0.37) and CS^b+ items (M = 0.84, SD = 0.38) compared with CS⁻ items (M = 0.65, SD = 0.42; vs CS^a+: t₍₁₂₁₎ = 4.81, p < 0.001, d = 0.51; vs CS^b+: for t₍₁₂₁₎ = 3.57, p = 0.001, d = 0.39; main effect CS category: F_(2,242) = 13.24, p < 0.001, partial η² = 0.041), while there was no reliable difference between CS^a+ and CS^b+ items (t₍₁₂₁₎ = 1.24, p = 0.218, d = 0.10; Fig. 3C,D).

When analyzing the hit rates and false alarm rates separately, results showed that participants performed well in the surprise recognition test on Day 2, as reflected in significantly higher hit rates (M = 0.59, SD = 0.33) than false alarm rates (M = 0.41, SD = 0.37; t₍₁₂₁₎ = 16.16, p < 0.001, d = 0.50). Hit rates differed significantly between CS conditions (F_(2,242) = 5.63, p = 0.004, partial η² = 0.005), as did false alarm rates (F_(2,242) = 6.47, p = 0.002, partial η² = 0.002). Post hoc comparisons revealed that the average hit rate for CS^a+ items (M = 0.62, SD = 0.13) was significantly higher than for CS^b+ items (M = 0.58, SD = 0.12; t₍₁₂₁₎ = 2.71, p = 0.007, d = 0.11) and CS⁻ items (M = 0.56, SD = 0.15; t₍₁₂₁₎ = 2.91, p = 0.004, d = 0.16), while the hit rates for CS^b+ and CS⁻ items did not differ (t₍₁₂₁₎ = 0.93, p = 0.355, d = 0.05). For the false alarm rate, there was a comparable pattern with the average false alarm rate being significantly increased for CS^a+ (M = 0.42, SD = 0.09) items compared with CS^b+ items (M = 0.39, SD = 0.09; t₍₁₂₁₎ = 2.48, p = 0.014, d = 0.08) but not to CS⁻ items (M = 0.43, SD = 0.08; t₍₁₂₁₎ = −0.66, p = 0.509, d = 0.02). Notably, the false alarm rate was significantly lower for CS^b+ items compared with CS⁻ items (t₍₁₂₁₎ = −3.65, p < 0.001, d = 0.10).

CTBS over the SPC had a significant impact on overall recognition memory performance. For d′, our analyses revealed significantly increased memory performance for participants of the TMS group (M = 0.87, SD = 0.53) compared to those of the sham group (M = 0.72, SD = 0.41; t_(116.2) = 2.71, p = 0.032, d = 0.39; Fig. 3C,D). To further elucidate where this difference in memory sensitivity is coming from, we analyzed the hit rates and false alarm rates for each experimental group. Participants who received cTBS immediately before the encoding session had a significantly lower hit rate (M = 0.38, SD = 0.20) but also a significantly lower false alarm rate (M = 0.17, SD = 0.22) compared with the participants of the sham group (hit rate: M = 0.81, SD = 0.09, t_(115.0) = −9.57, p < 0.001, d = 1.74; false alarm rate: M = 0.67, SD = 0.33, t_(99.3) = −10.13, p < 0.001, d = 1.85), suggesting that cTBS over the SPC led to more conservative mnemonic responses. For hit rates and false alarm rates in the TMS and sham groups per CS category, respectively, see Figure S1. However, the observed pattern of memory performance may have been influenced by the fact that we had half as many foils compared with targets, which could have affected participants’ response bias.

Signed PEs enhance memory for preceding stimuli

To investigate the influence of PEs on memory formation, we fitted GLMMs with recognition of an item as the binary dependent variable and added relevant independent predictors in a stepwise manner.

We started with a minimal model, in which we tested whether trial-wise sPEs contribute to item recognition. We treated the sPE (ranging from −1 to 1) after a CS item, i.e., the predictive stimulus, as the sole independent variable to predict subsequent item recognition. Estimates obtained revealed that sPEs, z = 2.44, p = 0.015, ß = 0.09, showed the expected positive relationship with subsequent item recognition. To rule out that the sPE effects were confounded with the explicit shock prediction or the outcome (shock vs no-shock), we computed a model in which we added the explicit shock prediction and the outcome, respectively, as additional predictors. Importantly, the memory-enhancing effect of the subsequent sPEs remained significant (all z > 3.95, p < 0.001, ß > 0.18; Fig. 3E) after accounting for prediction or the outcome, respectively. To further exclude that the PE effects were solely driven by shock omission, we ran a separate model within the CS^a+ and CS^b+ categories, which also revealed significant enhancing effects of sPEs on memory (z = 2.56, p = 0.011, ß = 0.09).

To investigate whether sPE effects on memory are dependent on the CS-outcome delay, we set up a model in which we included the sPE, the explicit shock prediction, the CS-outcome delay, and the sPE × CS-outcome delay interaction to predict the binary item recognition. This model revealed that memory was not influenced by the CS-outcome delay (z = 0.46, p = 0.648, ß = 0.00) or by the sPE × CS-outcome delay (z = 0.20, p = 0.841, ß = 0.00) but significantly enhanced by the sPE (z = 2.95, p = 0.003, ß = 0.19). These findings suggest that the memory-enhancing effect of sPEs is not affected by the delay between CS and outcome, in line with previous findings (Loock et al., 2025).

Next, we set up a model that incorporated the stimulation group (TMS vs sham), sPE, their interaction, and the shock prediction to examine whether cTBS over the SPC modulated the sPE effects on memory. Again, we found a memory-enhancing effect of sPEs, z = 3.42, p = 0.001, ß = 0.22. As expected, the stimulation group also affected item memory (z = −10.18, p < 0.001, ß = −2.75) suggesting that memory, expressed as hits, decreased when cTBS was applied. However, as the above analysis of overall memory performance showed, this effect was most likely due to a more conservative response bias in the TMS group. Importantly, the sPE × stimulation group interaction was not significant (z = −0.64, p = 0.525, ß = −0.05) suggesting that the sPE effect on memory was not modulated by cTBS over the SPC.

PEs trigger neural stimulus category reactivation

In line with previous findings (Loock et al., 2025), our behavioral data showed that the sPEs enhance subsequent memory for stimuli preceding the PE, irrespective of the time interval between the predictive stimulus and the outcome (i.e., PE). CTBS over the SPC appeared to not modulate the sPE effect on memory. In a next step, we investigated the neural responses elicited by a PE.

Event-related potentials

First, we analyzed electrophysiological responses to the outcome, i.e., shock versus no-shock, reflected in the FRN and P3. Therefore, we computed LMMs that included the sPE, the stimulation group, CS-outcome delay, sPE × CS-outcome delay interaction, sPE × stimulation group interaction, sPE × CS-outcome delay × stimulation group interaction, and the explicit shock prediction to predict the outcome-evoked FRN amplitude and P3 amplitude, respectively. For FRN amplitudes (0–800 ms after outcome onset at Fz), we found a significant effect of sPEs (t₍₁₂₉₈₎ = 2.68, p = 0.007, ß = 3.53), while there were no significant effects of the stimulation group (t_(119.6) = −0.97, p = 0.333, ß = −1.66), the CS-outcome delay (t_(337.5) = −1.83, p = 0.068, ß = −0.20), or the sPE × CS-outcome delay × stimulation group interaction (t_(40264.6) = 1.05, p = 0.293, ß = 0.31). As expected and in line with previous research showing a role of the FRN in error processing (Holroyd and Coles, 2002; Bellebaum and Daum, 2008), these findings show that the FRN is scaled by sPE magnitude, i.e., more pronounced with increasing sPEs, irrespective of CS-outcome delay or stimulation group.

For P3 amplitudes (300–1,000 ms after outcome onset at Pz), we found neither a significant effect of the sPE (t_(2406.4) = 1.21, p = 0.226, ß = 6.35) nor of any other predictor (all p > 0.222) suggesting that none of them affected the P3 significantly.

Time–frequency analyses

In a next step, we assessed whole-brain time–frequency patterns data after an outcome was revealed. We obtained a significant positive cluster emerging at ∼0.5 until 2 s after outcome onset at parieto-occipital electrodes in alpha and beta bands (9–18 Hz; electrode: PO8; p < 0.001, ci-range < 0.01, SD < 0.01). Additionally, we found a significant negative cluster emerging at ∼1.15 until 3 s after outcome onset at frontocentral electrodes in theta and alpha bands (7–13 Hz; electrode: C1; p < 0.001, ci-range < 0.01, SD < 0.01) suggesting that there was an early alpha synchronization and a late theta desynchronization after outcome onset (Fig. 4A).

Figure 4.

Outcome-evoked changes, PEs, and memory. A, Averaged time–frequency representations of channels C1 (top panel) and PO8 (bottom panel) showing significant clusters as revealed by cluster-based permutation tests (outlined in black). B, Overall, a GLMM showed a significant three-way interaction effect of theta power × signed PE × stimulation group to predict subsequent memory. Notably, the statistical model considered the signed PE (sPE) as a continuous variable ranging from −1 to 1 (sPE). For displaying purposes, the continuous sPE variable was split into negative (–1 to 0) and positive (0 to 1) PEs to plot separate regression lines. Signed PEs and theta power interactively predicted memory performance in the sham group (top panel), but not in the TMS group (bottom panel). Theta power was log-transformed. C, Trial-wise category pattern reactivation computed by multivariate pattern analysis of EEG data. An L2-penalized logistic classifier was trained to classify between neural patterns of three categories from the DMS task (scenes, animals, tools) and tested on the same categories in the incidental encoding-fear learning task (top panel). Overall classification accuracy was higher following negative PEs compared with positive PEs (bottom panel). All depicted images are licensed under Creative Commons BY-SA license.

Next, we investigated whether the oscillatory power in the alpha (8–12 Hz) and theta bands (4–7 Hz), respectively, in a 3 s outcome-evoked window scaled with PE magnitude. To this end, we computed LMMs that treated the sPE and the CS-outcome delay including their interaction as independent variables to predict oscillatory power in the alpha and theta band separately after the outcome was presented. For alpha, this analysis showed no significant effects of sPEs (t_(20055.8) = 0.30, p = 0.766, ß = 0.00), nor of CS-outcome delay (t_(118.6) = −0.34, p = 0.732, ß = 0.00) and no sPE × CS-outcome delay interaction (t_(36776.7) = 0.10, p = 0.922, ß = 0.00). Likewise, for theta, we obtained no significant effects of sPEs (t_(8408.4) = 0.47, p = 0.636, ß = 0.00), nor of CS-outcome delay (t_(138.4) = 0.34, p = 0.738, ß = 0.00) and sPE × CS-outcome delay interaction (t_(37119.8) = −0.43, p = 0.666, ß = 0.00). These results indicate that the oscillatory outcome-evoked power in the alpha and theta band seems to be unaffected by the sPE and by the interval between predictive stimulus and outcome.

To control for potential effects of anticipatory mechanisms during the CS presentation, we also investigated whether neural activity, i.e., N1 component, P3 component, and alpha and theta oscillations, during cue presentation interacted with subsequent PEs and their effects on memory but found no significant effects (Text S1).

Decoding of category representations

In a next step, we leveraged an EEG-based decoding approach to investigate the neural patterns after the occurrence of a PE, i.e., potential category reactivation. As expected, the average performance of participants in the DMS task (i.e., during classifier training) was very high (M = 95.14% correct, SD = 0.12) and did not reliably differ between stimulation groups (t_(70.7) = 1.84, p = 0.069, d = 0.38).

Overall, the decoding accuracy during the outcome-evoked window (averaged over participants' individual peak accuracies) was significantly above chance (M = 0.55, SD = 0.03, t₍₁₁₇₎ = 23.04, p < 0.001, d = 4.26; Fig. 4C) and significantly higher in the TMS group (M = 0.56, SD = 0.03) than in the sham group (M = 0.55, SD = 0.02; t₍₁₁₂₎ = 3.15, p = 0.002, d = 0.58).

Next, we investigated whether the PE magnitude was associated with category reactivation after the outcome by setting up a LMM that included the sPE to predict the decoding accuracy in the outcome-evoked window. Our analysis revealed a significant sPE effect on decoding accuracy (t_(117.1) = −2.43, p = 0.017, ß = −0.02). Interestingly, decoding accuracy was significantly higher following negative PEs (i.e., unexpected shock omissions, M = 0.57, SD = 0.63) compared with positive PEs (i.e., unexpected shock presentations, M = 0.56, SD = 0.64; z = 2.92, p = 0.004, ß = 0.02; Fig. 4C). This result indicates that outcome-evoked category reactivation was mainly driven by negative sPE.

To investigate whether cTBS affected the PE-related category reactivation, we set up a LMM that included the sPE × stimulation group interaction. Importantly, there was no significant sPE × stimulation group interaction (t_(115.8) = −1.48, p = 0.143, ß = −0.02) on outcome-evoked decoding accuracy, suggesting that the sPE effect on outcome-evoked neural category reactivation seems to be unaffected by cTBS over the superior parietal lobe.

Outcome-evoked theta power boost predicts item recognition after negative PEs

In order to test whether outcome-evoked neural changes predicted subsequent memory, we first computed a GLMMs that tested whether the FRN amplitude, in response to a PE, or the FRN amplitude × sPE interaction predicted the binary item recognition. However, we found no significant effect of the FRN amplitude, z = 1.22, p = 0.224, ß = 0.00, nor a significant FRN amplitude × sPE interaction, z = −1.16, p = 0.248, ß = −0.00, on subsequent item memory.

Then, we set up GLMMs that included the outcome-evoked oscillatory spectral power in the alpha and theta bands, respectively, sPE and stimulation group including their interaction to predict item recognition. When considering alpha power, we obtained no significant sPE × average spectral power × stimulation group interaction (z = 1.25, p = 0.212, ß = 0.30) on subsequent memory. Interestingly, when considering theta power in a separate GLMM, we found a significant sPE × average spectral power × stimulation group interaction (z = 2.96, p = 0.003, ß = 0.68) on memory. We pursued this effect with follow-up GLMMs for the sham and TMS group separately. In the sham group, we found a significant sPE × theta band power interaction (z = −2.65, p = 0.008, ß = −0.51), whereas there was no such effect in the TMS group (z = 1.33, p = 0.185, ß = 0.17; Fig. 4B). Although theta power was unaffected by the PE per se, our analysis showed that, in the sham group, the relationship between the outcome-evoked theta power and subsequent recognition memory appeared to depend on the nature of the PE. We pursued this effect in the sham group with GLMMs in which we added the sPE × theta power interaction as a variable to predict item recognition separately for datasets only containing negative (sPE < 0) and positive PEs (sPE > 0). Specifically, following negative PEs, increased theta power was significantly associated with enhanced item recognition (z = 2.12, p = 0.029, ß = 0.30), whereas for positive PEs, theta power was not significantly associated with item recognition (z = 0.43, p = 0.670, ß = 0.06). CTBS over the SPC appeared to abolish this association.

Next, we investigated whether the PE-induced changes in the category reactivation in the outcome-evoked window, as assessed by MVPA-based decoding, predicted subsequent item recognition. A minimal GLMM in which we included outcome-evoked decoding accuracy to predict item recognition revealed no significant effect of decoding accuracy (z = 0.81, p = 0.416, ß = 0.03). In a follow-up GLMM, we included the sPE, stimulation group, and the sPE × decoding accuracy × stimulation group interaction as additional predictors in the previous model. Our analyses showed no significant interaction effects of sPE × decoding accuracy (z = 0.07, p = 0.942, ß = 0.01) and sPE × decoding accuracy × stimulation group (z = −0.38, p = 0.701, ß = −0.05) on item recognition.

PE effect on subsequent memory depends on the neural state shortly before the PE

In a next step, we tested whether PE effects on subsequent memory depend on the neural state and potential stimulus category representation shortly before the PE.

Time–frequency analyses

First, we investigated whether the average spectral power in the 3 s before the outcome interacted with the sPE on subsequent memory. Critically, in this analysis, we disregarded all trials in which the analyzed window could have overlapped with the presentation of the predictive stimulus to ensure temporal separation, i.e., excluding all trials in which the CS-outcome delay was shorter than 3 s. Notably, when including all trials, irrespective of the CS-outcome delay duration, we observed a very similar pattern of results.

We set up GLMMs that included pre-outcome alpha power (8–12 Hz) and theta power (4–7 Hz), respectively, sPE and stimulation group including their interaction to predict item recognition. When including alpha power, our analysis yielded a significant sPE × alpha power × stimulation group interaction (z = 2.55, p = 0.011, ß = 0.84) on memory. We pursued this effect with separate follow-up GLMMs for the sham and TMS group, respectively. In the sham group, we found a significant sPE × alpha power interaction (z = −2.95, p = 0.003, ß = −0.85), while there was no such interaction effect in the TMS group (z = 0.06, p = 0.949, ß = 0.01; Fig. 5). Similarly, when including theta power, we obtained a significant sPE × theta power × stimulation group interaction (z = 3.18, p = 0.001, ß = 1.03) on memory. Again, we pursued this effect with separate follow-up GLMMs for the sham and TMS group, respectively. In the sham group, we found a significant sPE × theta power interaction (z = −2.92, p = 0.004, ß = −0.79), which was not present in the TMS group (z = 1.45, p = 0.148, ß = 0.26; Fig. 5). These findings suggest that the impact of sPE on item recognition was modulated by the pre-outcome neural oscillatory activity in the sham group. Again, we pursued this effect in the sham group with GLMMs in which we added the sPE × alpha power interaction and sPE × theta power interaction, respectively, as a variable to predict item recognition separately for datasets only containing negative and positive PEs. Specifically, before negative PEs, increased alpha and theta power were significantly associated with enhanced item recognition (alpha: z = 2.30, p = 0.021, ß = 0.39; theta: z = 2.82, p = 0.005, ß = 0.45), whereas for positive PEs, alpha power and theta power (alpha: z = −0.94, p = 0.345, ß = −0.17; theta: z = 0.46, p = 0.645, ß = 0.08) were not significantly associated with item recognition. CTBS over the SPC appeared to abolish these associations.

Figure 5.

Pre-outcome changes, PEs and memory. Overall, GLMMs showed significant three-way interactions of alpha power × signed PE × stimulation group (A) and theta power × signed PE × stimulation group (B), respectively, to predict subsequent memory. Notably, all statistical models considered the signed PE (sPE) as a continuous variable ranging from −1 to 1. For displaying purposes, the continuous sPE variable was split into negative (–1 to 0) and positive (0 to 1) PEs to plot separate regression lines. Alpha and theta power were log-transformed. A, sPEs and alpha power interactively predicted item memory in the sham group (top panel), but not in the TMS group (bottom panel). Shaded areas represent 95% confidence intervals. B, A similar pattern was obtained for theta power, with a significant sPE × theta power interaction on item memory in the sham group (top panel) but not in the sham group (bottom panel). C, Stronger stimulus (category) representations above chance level (dotted line) shortly before the PE impair memory in the context of positive PEs. Slope truncations reflect data sparsity at extreme ends of the predictor variables.

Decoding of stimulus category reactivation

Next, we leveraged an EEG-based decoding approach again to investigate the neural patterns shortly before the occurrence of a PE, i.e., potential stimulus (category) representation. Same as in the time–frequency analysis, we disregarded all trials in which the CS-outcome delay was shorter than 2 s to make sure that the analyzed window did not overlap with the stimulus presentation.

Overall, the decoding accuracy during the pre-outcome window (averaged over participants’ individual peak accuracies) was significantly above chance (M = 0.55, SD = 0.03, t₍₁₁₇₎ = 22.84, p < 0.001, d = 4.22; Fig. 5C) and significantly higher in the TMS group (M = 0.57, SD = 0.03) than in the sham group (M = 0.55, SD = 0.03; t_(115.7) = 4.36, p < 0.001, d = 0.80). A LMM including the stimulation group and CS-outcome delay to predict pre-outcome decoding accuracy on trial level confirmed that there was a significant effect of stimulation group (t_(116.4) = 3.77, p < 0.001, ß = 0.16, d = 0.70) but no significant effect of CS-outcome delay (t_(33663.6) = 1.15, p = 0.251, ß = 0.00, d = 0.01).

Next, we set up a GLMM in which we included the sPE and the sPE × decoding accuracy interaction to predict subsequent item recognition. This analysis yielded a significant effect of sPEs (z = 3.31, p = 0.001, ß = 0.16) and a significant sPE × decoding accuracy interaction effect (z = −2.10, p = 0.035, ß = −0.14) on item memory suggesting that the PE effect on memory performance depends on the strength of the stimulus (category) maintenance shortly before the PE (Fig. 5C). Again, we pursued this effect with GLMMs in which we added the sPE × pre-outcome decoding accuracy as a variable to predict item recognition separately for datasets only containing negative and positive PEs. For negative PEs, pre-outcome decoding accuracy was not significantly associated with subsequent item recognition (z = 1.24, p = 0.216, β = 0.06), whereas for positive PEs, increased decoding accuracy was associated with decreased memory performance (z = −2.59, p = 0.010, β = −0.13). These findings suggest that stronger stimulus (category) representation shortly before the PE seems to impair memory encoding in the context of positive PEs, whereas this effect is absent for negative PEs. A follow-up model incorporating stimulation group yielded that the PE × decoding accuracy was not affected by cTBS over the right SPC, i.e., a nonsignificant effect of stimulation group × decoding accuracy (z = 0.01, p = 0.938, ß = 0.01) and a nonsignificant sPE × decoding accuracy × stimulation group interaction on item recognition (z = −0.66, p = 0.511, ß = −0.10).

Additionally, to investigate the role of uPEs in memory formation, we first computed all models replacing sPEs with uPEs (Table S1). Across all models, uPEs did not show any significant main or interaction effects on memory performance (all p > 0.162). Model comparisons between uPE and sPE models—both pre-outcome and outcome-evoked—based on AIC and log-likelihood indicated that uPE models provided equivalent fits for most models. Notably, uPE models showed a better fit when predicting general memory performance compared with the corresponding sPE models (Table S1). To further disentangle the unique contributions of sPEs and uPEs, we computed a combined model including both as predictors—along with the explicit shock prediction—for item recognition. This combined model demonstrated an improved fit (AIC = 35,635; log-likelihood = −17,804) relative to models with sPE or uPE alone (Table S1). Critically, sPEs significantly predicted memory performance in this model (z = 3.21, p = 0.001, ß = 0.15), whereas uPEs did not (z = −0.54, p = 0.593, ß = −0.05), indicating that the memory-enhancing effect in our paradigm is specifically driven by sPEs.

Control variables

Importantly, participants were not aware of their actual stimulation condition (sham vs TMS) as assessed by a treatment guess at the end of the experiment (χ²₍₁₎ = 0.278; p = 0.599). Moreover, participants were moderately surprised by the recognition test on Day 2 (M = 3.13, SD = 1.27). Participants who underwent the cTBS appeared to be less surprised by the recognition test (M = 2.42, SD = 1.18) than participants of the sham group (M = 3.83, SD = 0.92; t_(109.88) = 7.26, p < 0.001, d = 1.39). Seventeen participants chose the “not surprised at all” option. Because excluding them did not affect the results, we included them in all analyses. Additionally, we included participants’ surprise into the model to analyze overall memory performance, i.e., d’, but the group difference remained (F_(1,115) = 5.87, p = 0.017, partial η² = 0.049), suggesting that it cannot be explained by the difference in the surprise related to the recognition test.

State and trait anxiety levels were comparable between stimulation groups (STAI-T: t₍₁₁₈₎ = 0.08, p = 0.935, d = 0.02; STAI-S: t₍₁₁₈₎ = 0.95, p = 0.346, d = 0.18). Notably, there were significant differences between groups in depressive mood (BDI-II: t₍₁₁₈₎ = 6.12, p < 0.001, d = 1.13), chronic stress (TICS: t₍₁₁₄₎ = 5.41, p < 0.001, d = 1.01), and sleep quality (t₍₈₆₎ = 5.55, p < 0.001, d = 1.20) suggesting that those characteristics were increased in the TMS group before cTBS was applied. Importantly, adding those measures as covariates in our analyses did not change the pattern of results, suggesting that these variables could not explain the group differences we obtained.

Discussion

PEs play a pivotal role in adaptive memory formation (Kalbe and Schwabe, 2020, 2022; Rouhani et al., 2023; Loock et al., 2025), yet the underlying mechanisms of these effects remain unclear. Employing EEG and MVPA to investigate the neural dynamics surrounding PEs, our results demonstrate that the impact of the PEs on memory formation is driven by neural states immediately before the PE and outcome-evoked neural changes.

Our results replicate the previously reported beneficial effect of PEs on episodic memory for preceding stimuli. Previous studies have often examined uPEs (Kalbe and Schwabe, 2020), reflecting the overall PE magnitude regardless of its direction. However, recent evidence suggests that positive and negative PEs differentially affect memory formation (Rouhani and Niv, 2021; Kalbe and Schwabe, 2022; Loock et al., 2025). Therefore, we focused on continuous sPEs to distinguish between positive and negative PEs. Our behavioral data corroborate previous findings (Loock et al., 2025) showing that negative PEs were associated with impaired subsequent memory, whereas positive PEs were linked to memory enhancement. No such effects were obtained for uPEs.

While initial fMRI evidence suggests that PE-induced effects on memory are associated with decreased activation of the medial temporal lobe and enhanced interaction between the salience and a frontoparietal network (Kalbe and Schwabe, 2022), we focused on neural processes surrounding a PE event. Importantly, these processes differ from schema-related effects (van Kesteren et al., 2012), typically emerging during encoding when schemas are already established. In contrast, in our paradigm, PEs arise after encoding and retroactively influence memory formation for predictive, but not for uninformative stimuli (Loock et al., 2025). Although prestimulus (Cohen et al., 2014) and outcome-evoked reactivation (Carr et al., 2011; Deuker et al., 2013) are well known to be linked to subsequent memory, the memory enhancements observed here depends on neural dynamics surrounding the PE. These dynamics interact with PEs, indicating that the memory benefits are contingent on the PE rather than generally elevated pre- or poststimulus activity. Notably, our decoding results are limited to category-level reactivation. Although we cannot distinguish whether this reflects reactivation or maintenance pre-outcome, both processes mechanistically serve to keep relevant information accessible, thus allowing an interaction with PEs. We interpret the outcome-evoked decoding as reactivation, triggered by the outcome, consistent with PE-driven memory updating (Sinclair et al., 2021).

We predicted that the sPE effects on memory may require neural representations of the predictive stimulus shortly before the PE event. Remarkably, our findings provide evidence that the sPE effect on memory depends on the strength of the category representation shortly before the PE. Interestingly, our findings show for positive PEs, i.e., unexpected shock presentations, that increased stimulus category representation before the PE was associated with impaired memory performance. This result could suggest that strong reactivation of the predictive stimulus category increases interference with the attention-grabbing unexpected shock, thereby impairing memory storage. In contrast, negative PEs, i.e., unexpected shock omissions, did not show this effect, presumably because they are less emotionally salient.

Beyond the neural category reactivation, we also analyzed the impact of neural oscillations before the PE event on PE-related memory effects. Interestingly, both alpha and theta activity before the PE were associated with sPE-induced memory changes (in the sham group). These effects were dependent on the sign of the PE and distinguishable from mere neural activity induced by the predictive stimulus. For negative PEs, increased pre-outcome alpha and theta activity were linked to improved subsequent memory, potentially due to enhanced attentional processing and associative binding (Staudigl and Hanslmayr, 2013; Payne and Sekuler, 2014). For positive PEs, however, alpha and theta activity showed no significant relationship with memory, presumably due to increased interference from the unexpected shock disrupting encoding processes.

In addition to the relevance of the neural state shortly before the PE, we proposed an alternative mechanism suggesting that PEs may induce neural changes that facilitate the memory storage of preceding stimuli. As expected, sPEs led to an increased FRN, consistent with findings linking the FRN with error processing (Holroyd and Coles, 2002; Bellebaum and Daum, 2008). Intriguingly, sPEs were also associated with increased category reactivation. We found increased category reactivation after negative PEs compared with positive PEs, suggesting that negative PEs elicit stronger category reactivations after the outcome, presumably reflecting an adaptive mechanism for updating predictive models in response to unexpected safety signals, i.e., unexpected shock omissions. This aligns with research demonstrating that unexpected omissions of aversive events engage mnemonic processes related to model updating (Iglesias et al., 2013), also suggesting that the increased category reactivation following negative PEs may reflect increased attentional processing due to the unexpected relief from an anticipated aversive event (Li et al., 2011; Kalbe and Schwabe, 2022). In contrast, positive PEs, representing an unexpected aversive outcome, presumably impair category reactivation due to a more pronounced processing of the unexpected shock, presumably leading to arousal-biased competition of attention where attentional resources are preferentially allocated to the aversive event rather than to the preceding predictive stimulus (Mather and Sutherland, 2011; Kalbe and Schwabe, 2022). Surprisingly, category reactivation after a PE was not related to item recognition suggesting memory formation does not depend on outcome-evoked reactivation. Category reactivation may primarily support model updating rather than memory consolidation (Iglesias et al., 2013). Particularly in the context of aversive learning, neural responses after outcome representation may reflect an adaptive updating of future expectations rather than strengthening individual item memory (Schwiedrzik and Freiwald, 2017). Furthermore, physiological arousal induced by the PE may interact with category reactivation, such that the relevance of stimulus information depends on whether attentional resources are directed toward updating predictive models rather than encoding specific stimulus details (Mather and Sutherland, 2011).

Although theta and alpha power were not directly modulated by sPEs, our findings indicate an interplay of sPEs and outcome-evoked theta power on recognition, depending on the sign of the PE. Following negative PEs, theta oscillations were associated with better item recognition indicating that theta waves support memory consolidation when an expected aversive event does not occur. This could reflect enhanced memory storage under conditions of surprise by attenuating processes in the default mode network which may otherwise impair memory formation (Klimesch, 1999; White et al., 2013; Kota et al., 2020), which might be due to theta oscillation-induced binding of an item to its spatiotemporal context in the medial temporal lobe and in hippocampo-cortical feedback loops (Klimesch, 1999; Hanslmayr et al., 2011). For positive PEs, there was no such relationship indicating that theta oscillations may be more relevant for encoding unexpected safety signals rather than unexpected threat.

We also investigated whether and how the PE effects on memory and their underlying neural mechanisms could be modulated using brain stimulation. We focused on the SPC, given its critical role in working memory processes and top-down attentional updating (Corbetta et al., 1995; Koenigs et al., 2009; D’Esposito and Postle, 2015), and applied cTBS over the right SPC to modulate PE effects on memory. Overall, we found an increase in memory performance in the cTBS group primarily driven by a more conservative response bias. This finding resonates with the observed increase in neural category reactivation of the predictive stimulus following cTBS. Inhibiting the SPC may have reduced competing attentional processes, thereby facilitating more targeted reactivation of relevant memory traces. Given its role in top-down attentional control (D’Esposito and Postle, 2015), an inhibition might have reduced interference during retrieval and enhanced task-relevant category reactivation, i.e., stronger decoding signals, consistent with findings that cortical reinstatement enhances memory retrieval by reactivating content-specific encoding patterns (Gordon et al., 2014). Importantly, the behavioral PE effect on memory was not significantly altered by cTBS suggesting that intact functioning of the SPC is not essential for the PE effects on general memory performance. Instead, the beneficial effect of PEs on memory might derive from more distributed neural mechanisms, potentially involving medial temporal and prefrontal regions (Kalbe and Schwabe, 2022) involved in the detection of expectancy violations and the prioritization of salient event information for long-term storage. However, at the neural level, the mechanisms underlying the PE effect on memory were altered by cTBS. Specifically, the neural signatures of PE-related memory effects, which were evident in the sham group, were diminished in the cTBS group. The SPC is assumed to be critical for pre-outcome neural states that then shape PE-related memory effects, presumably by facilitating attentional and predictive processes (Cabeza et al., 2008). However, the absence of this effect in the TMS group suggests that cTBS may have disrupted the anticipatory modulation of encoding processes by pre-outcome alpha and theta activity, thereby weakening the interplay between preparatory neural states and PEs.

In summary, we demonstrate that the effects of sPE on memory formation for preceding events are influenced by neural states and representations surrounding the PE. Specifically, the impact of PEs on memory depends on theta and alpha oscillations shortly before the PE occurs, which may provide attentional and mnemonic binding resources required for PE-induced modulation of memory formation. More generally, these results contribute to our understanding of mechanisms underlying adaptive memory, where stimuli predicting emotionally relevant events are particularly well stored in long-term memory.