What makes a scent trigger a memory? A cognitive decomposition of odor-evoked retrieval

Abstract

A single scent can unlock vivid memories. This study investigates the factors that make some odors more evocative than others. We examined odor-evoked episodic memory in 106 participants who experienced odors embedded in distinct visuospatial contexts, and whose memory was tested 24-72 h later. The protocol empirically dissociates odor recognition ("I've already smelled this scent") and associative memory ("It evokes a memory") processes. Using machine learning with SHapley Additive exPlanations, we identified distinct predictors for each process. Recognition was driven by emotional strength, especially for unpleasant odors, and the richness of verbal descriptions. Associative memory followed a U-shaped relationship with familiarity and was strongly influenced by semantic distinctiveness-how uniquely each odor was described. Together, these findings reveal that odor memorability depends not only on its emotional salience but also on how specifically it is conceptualized and how familiar we are with it.

Keywords: Cognitive neuroscience; Machine learning; Neuroscience.

Graphical abstract

Figure 1

Experimental task design (A) Time course of encoding sessions. Participants experienced one episode per day over 3 days. Each episode consisted of a landscape picture in which, while gray circles were inactive, each orange circle symbolized spatial location of a particular odor which was delivered when the circle was clicked. (B) Example of a retrieval trial, on day 4. Each trial began with an odor recognition task. Upon answering “Yes” to recognizing an odor, participants proceeded to the associative part of the episodic memory retrieval task and described the context in which this odor had been previously encountered, by first selecting a context (from the three pictures) and then selecting a location (among the nine circles).

Figure 2

Participants represent odors in distinct perceptual spaces Perceptual spaces of three participants computed (A) from all target odors for the recognition component and (B) from all hit odors for the associative component of odor-evoked memory. Each odor, evaluated by the participants in terms of pleasantness, familiarity, and intensity, is represented by a point in a three-dimensional space. The convex envelope (enveloping volume) around the points indicates the distribution of odors in each participant’s perceptual space. The size and shape of the envelope reflect the variability of participants’ perceptual evaluations of odors.

Figure 3

Errors in odor recognition reflect feature characteristics (A) AUC scores of different models during training and testing. (B) Confusion matrix showing the distribution of actual versus predicted odor recognition outcomes on the test set for both the base model and the final model. The final model improved classification performance for class 0. (C–H) Scatterplots showing individual test trials according to odor-related features: (C) intensity, (D) emotional strength, (E) pleasantness, (F) familiarity, (G) number of words in the description, (H) gender distribution per trial category, and (I) semantic distance between odors. (J) Dimensionality reduction of the seven features using t-SNE. These scatterplots illustrate the feature value ranges and their overlap between correctly and incorrectly classified trials: the feature value ranges of hits incorrectly identified as misses is close to those of correctly identified misses and the feature value ranges of misses incorrectly identified as hits is close to those of correctly identified hits. Correct class 0: correctly identified as miss (n = 17 trials); correct class 1: correctly identified as hit (n = 111); misclassified 0–>1: miss incorrectly identified as hit (n = 26); misclassified 1–>0: hit incorrectly identified as miss (n = 44).

Figure 4

Odor recognition is primarily driven by emotional intensity and verbal richness (A) SHAP summary plot showing the feature contributions to odor recognition predictions. Each point represents an SHAP value for a single trial, with color indicating the feature’s value (blue for low and pink for high). For the gender feature, blue corresponds to female and pink to male. Positive SHAP values shift the prediction toward correct recognition, while negative SHAP toward incorrect recognition. (B–H) SHAP dependence plots illustrating the influence of individual features on the model’s predictions: (B) number of words in the description, (C) emotional strength, (D) intensity, (E) pleasantness, (F) semantic distance between odors, (G) familiarity, and (H) gender. Each plot shows how the value of a single feature affects the prediction of odor recognition. (I) SHAP waterfall plot illustrating the contribution of features to an individual odor recognition prediction. For this trial, the participant described the odor in 6 words, rated its pleasantness at −3.53, intensity at 7.6, and familiarity at 1.8. The average semantic distance between this odor and the others was 0.97. The participant was female (gender = 0). The horizontal pink and blue arrows represent the contribution of each feature, shifting the model’s mean prediction (E[f(x)] = 0.5) to the specific prediction for this trial (f(x) = 0.658).

Figure 5

Errors in associative memory reflect feature characteristics (A) AUC scores on training and testing sets across all models. (B) Confusion matrices showing the distribution of actual versus predicted outcomes for the associative component, for both the test set of the base model and the final model. (C and D) Scatterplots showing individual test trials according to: (C) semantic distance between odors and (D) odor familiarity. (E) The feature value ranges of correct predictions misclassified as incorrect is similar to that of trials correctly identified as incorrect, while the distribution feature value ranges of incorrect predictions misclassified as correct is close to that of trials correctly identified as correct. Correct class 0: correctly classified as incorrect associative retrieval (n = 32), correct class 1: correctly classified as correct associative retrieval (n = 49), misclassified 0–>1: incorrect associative retrieval misclassified as correct associative retrieval (n = 35), and misclassified 1–>0: correct associative retrieval misclassified as incorrect associative retrieval (n = 32).

Figure 6

Odor associative memory is primarily driven by odor familiarity and semantic distinctiveness (A) SHAP summary plot showing the contributions of features to associative memory predictions. Each point represents an SHAP value for a single trial, with color indicating the feature’s value (blue for low, pink for high). Positive SHAP values shift the prediction toward correct associative retrieval, while negative values shift the prediction toward incorrect associative retrieval. (B and C) SHAP dependence plots illustrating the effects of (B) semantic distance between odors and (C) odor familiarity on the model’s predictions. The SHAP values, plotted against feature values, show how these features influence the model’s output for predicting associative retrieval, or its absence. (D) SHAP waterfall plot depicting the contribution of features to an individual associative memory prediction. For this trial, the participant rated the odor with a familiarity score of 8.6, and the average semantic distance between this odor and the others was 1. The horizontal pink bars represent the contribution of each feature in shifting the model’s mean prediction (E[f(x)] = 0.502) to the specific prediction for this trial (f(x) = 0.667).