Design for EEG experiment and neural decoding for memory reactivation.

(A) In the EEG experiment, participants performed an inference task in which they were asked to infer the correct answer-C based on the presented cue-A, using pretrained AB and BC associations. Element B, the “bridge information,” could be either a face or a building image. Both the cue-A and the answer-C were pictures of daily-life objects. EEG data were recorded during both the inference and an independent localizer phase. During the localizer phase, participants actively viewed a different set of face and building images. A classifier was trained on category-specific neural response patterns from the localizer phase and then applied to the inference phase. This approach enabled time-series decoding analyses to detect, within individual trials, whether category information of the bridge element-B was reactivated at any moment. (B) Using neural decoding to capture the memory reactivation of bridge information-B, we examined whether this neural signature was stronger during successful inference trials (left panel) and weaker under acute stress (middle panel). Additionally, we assessed whether the neural signature could be protected through retrieval practice (right panel).

Task Overview.

During the encoding phase, participants learned two types of associations: AB and BC, where each association consisted of pairs of images, forming interconnected triads (ABC). During the inference phase, picture A was presented as a cue to recall picture C. Importantly, participants were not informed about the inference requirement during the AB and BC encoding phases. The task was divided into two sessions, separated by a 24-hour interval. On Day 1, the encoding phase lasted approximately 15 minutes, during which two pictures were presented side-by-side on the screen. On Day 2, participants underwent either a stress (i.e., Trier Social Stress Test [TSST]) or a control manipulation, and then finish the memory tasks. Acute stress responses were assessed using heart rate measurements and subjective emotion ratings (i.e., Positive and Negative Affect Schedule [PANAS]). EEG data were recorded during both the inference and localizer phases, with the inference phase being the critical period of our analysis. During the inference phase, Cue-A was presented alongside the correct answer-C and two lure images from other learned ABC triads. The original AB and BC associations were also tested to enable subsequent trial-by-trial analyses of memory inference. Experiment1 and Experiment2 used the same task procedure, with experiment2 included the additional retrieval practice session in Day1, immedialy after the encoding (See Figure S1 for the procedure of Retrieval Practice phase).

Manipulation Checks.

(A) Acute stress induction significantly increased heart rate (HR), measured in beats per minute (BPM), during the Trier Social Stress Test (TSST). The heart rate response was defined as the mean BPM difference between the TSST and baseline phases, with the stress group showing a mean increase of 11.9 BPM (SD = 4.95) and the control group showing a mean increase of 1.7 BPM (SD = 3.98). (B) Stress significantly increased subjective ratings of negative valence. (C) Stress significantly decreased subjective ratings of positive valence. Note: Results from Experiment 1 are presented here. Similar stress responses were observed in Experiment 2, which are shown in Figure S2.

Memory inference performance.

(A) Memory inference accuracy across groups in Experiment 1 (Control-E1, Stress-E1) and Experiment 2 (Control-E2, Stress-E2). Acute stress significantly impaired inference accuracy in Experiment 1, whereas this impairment was mitigated by retrieval practice in Experiment 2. (B) Response times (RTs) during the memory inference test. In Experiment 1, participants under acute stress exhibited significantly longer RTs compared to controls. This stress-induced slowing was eliminated in Experiment 2 following retrieval practice. (C) Comparison of RTs between successful (S) and unsuccessful (US) inference trials. Successful inferences were associated with significantly faster RTs compared to unsuccessful ones. (D) Density plots illustrating the distribution of RTs for successful (-S) and unsuccessful (-US) trials across different experimental conditions. Vertical lines indicate the mean of RTs.

Bridge information were reactivated to support subsequent successful inference.

(A) Hypothesized reactivation of bridge information during the inference phase. To select the correct image-C based on cue-A, participants were hypothesized to mentally retrieve the bridge information-B (e.g., a face in the example trial, in orange). In some cases, the incorrect image (e.g., a building, in purple) might also be reactivated, or no image might be recalled during this phase. (B) Definition of successful and unsuccessful inference trials. In successful trials, both associations (A–B and B–C) were remembered, allowing for the correct inference of the A–C relationship (Orange Box). In contrast, in unsuccessful trials, despite remembering A–B and B–C, participants failed to correctly infer the A–C relationship (Purple Box). Trials in which either A–B or B–C associations were not encoded prior to inference were excluded from the analysis (White Box). (C) Results of time-generalization analysis from within-task classification. EEG data from the localizer phase were used to train a machine learning model to classify faces and buildings based on localizer-phase EEG activity. (D) Results of time-generalization analysis from cross-task classification. EEG data from the localizer phase were used to train a machine learning model to classify faces and buildings using EEG data from the inference phase. Decoding accuracy matrices were computed separately for successful and unsuccessful inference trials (depicted in orange and purple, respectively). (E) Mean decoding accuracy within the optimal decoding time window, relative to chance, for successful and unsuccessful inference trials. Significantly above-chance decoding of bridge information predicted subsequent successful inference, demonstrating the Subsequent Inference Effect.

Neural reactivation under acute stress and after retrieval practice.

Decodability (A) and Mean Decoding Accuracy (B) of the bridge element during the Inference Stage. Acute stress reduced neural reactivation in Experiment 1 (No-Retrieval Practice) but enhanced it in Experiment 2 (Retrieval Practice). Decodability (C) and Mean Decoding Accuracy (D) of the bridge element during the Choice Stage. In contrast to the inference stage, neural reactivation during the choice phase was not significantly modulated by acute stress or retrieval practice, as indicated by the non-significant interactions. (E) Schematic of the trial structure, illustrating the distinct inference and choice phases analyzed. These findings indicate that the modulatory effects of stress and retrieval practice on bridge element reactivation are confined to the inference process itself.

Acute stress reduces theta power during memory inference, an effect counteracted by retrieval practice.

(A) Time-resolved difference waves (Control group – Stress group) of theta power for Experiment 1 (E1, blue; without retrieval practice) and Experiment 2 (E2, orange; with retrieval practice). The thick horizontal bars denote significant time clusters where theta power was greater in the control group than in the stress group (cluster-based permutation test, p < 0.05). (B) Time-frequency map of the t-statistic for the control vs. stress group contrast in E1, covering 2–12 Hz from 0 to 4 s. (C) The difference wave for the direct comparison between experiments (E2 – E1; magenta line) shows that retrieval practice significantly increased theta power. The thick horizontal bar highlights the significant temporal cluster. (D) Corresponding time-frequency map of the t-statistic for the E2 vs. E1 contrast. In all panels, time 0 s indicates the onset of the bridge element B cue. The dashed vertical lines delineate the 4-s analysis window for the memory inference phase. The colorbar is applicable to both time-frequency maps.