Prediction error and memory updating: Challenges ahead

Our ability to predict future events and experiences is essential in guiding goal-directed behaviors (Bar, 2009). Such predictions are based on our prior experiences that shape our semantic memory, i.e., knowledge of facts, and episodic memory, i.e., detailed representations of past events. When there are inconsistencies between predictions and subsequent experiences, incoming new information from ongoing experiences can result in updating of predictions to better match experiences (Niv & Schoenbaum, 2008). This short commentary is centered on papers from three research teams that are featured in a Special Issue section titled “From Memory to Predictions and Back: How Memories Form Predictions that Shape Memories.” These papers present systematic reviews exploring the role of prediction error across various aspects of learning and memory, including the fundamental mechanisms of neural integration and segregation, event segmentation in episodic memory, and developmental changes in episodic memory.

Bein and colleagues (Bein et al., 2023) conducted a comprehensive review of recent studies exploring how the accuracy of the brain’s memory-based predictions about upcoming events leads to neural integration and separation during the formation of semantic memory. They highlight studies employing innovative designs and advanced multivariate functional MRI analyses, concluding that accurate predictions facilitate neural integration of related memories, while prediction errors drive both integration and separation. Additionally, the authors emphasize the importance of identifying factors that influence the interaction between predictions and memory, such as memory reactivation, the strength of the prediction error, and the specific task designs. Nolden and colleagues (Nolden et al., 2024) conducted a comprehensive review of studies on the interplay between prediction error and event segmentation, a process that divides the continuous stream of experiences into discrete events, which is crucial for the formation of episodic memories. A key contribution of this paper is its effort to foster discussion on the shared mechanisms underlying memory formation through prediction error and event segmentation, with a particular focus on the roles of attention and working memory processes. Shing and colleagues (Shing et al., 2023) focused on how prediction error influences the neurocognitive mechanisms supporting episodic and semantic memory formation across the lifespan. They suggest that prediction error may promote synaptic changes, enhancing encoding and consolidation processes that likely have a greater impact on episodic memory, which relies on updates from sensory input, than semantic memory, which is based on pre-existing knowledge. They further proposed that prediction error effects differ significantly across the lifespan. Children may generate weaker predictions due to limited knowledge that would trigger greater plasticity for updating their knowledge, whereas adults with established knowledge may generate more refined predictions but demonstrate limited effects of prediction error to update knowledge.

While the three papers provide compelling overview of interesting research about the interactions between prediction error and memory updating, they also highlight certain gaps that merit further exploration. Although empirical efforts have sought to understand the neurocognitive and neural mechanisms underlying prediction errors (e.g., Bein et al., 2020; Barron et al., 2020; Sinclair et al., 2021; Köster, 2024), an integrated framework explaining how these errors initiate a cascade of brain processes across multiple regions, ultimately enhancing episodic memory formation remains elusive. Other gaps that are identified center around the role of context, uncertainty, and reward prediction in modulating prediction errors. Future research should address these gaps to contribute to more refined models that include neurobiological and individual factors that influence prediction error interactions with memory.

Perhaps the best place to start in attempting an integration across the three papers that compose this section is by carefully examining the way each team defines prediction error. Such examination underscores a need for a comprehensive conceptualization of prediction error without which it is challenging to make foundational advance in understanding the role of prediction error in memory updating and its neural foundations. Bein and colleagues (Bein et al., 2023) define prediction error as the difference between one’s expectations based on existing knowledge and the “evidence,” which reflects what subsequently occurred. Nolden and colleagues (Nolden et al., 2024) define prediction error as the disparity between predictions and actual experiences. Shing and colleagues (Shing et al., 2023) offer a more specific definition, focusing on mnemonic prediction error, which commonly arises in daily experiences when subsequent experiences deviate from predictions based on contextual knowledge and memory. Any difference between those definitions likely results in differences in the scope of the phenomena described, and the neural basis that explains them. Integration across papers would be most successful if they define the central concept of prediction error more similarly.

One way to test the robustness of capturing the key features of prediction error influence on memory updating may involve adding a theoretically independent dimension, such as valence, and examine how variance in this dimension influences the effect of prediction error on memory updating. Notably, those definitions leave out considerations of valence prediction error — those related to reward or punishment—arguing that the underlying cognitive processes may differ, and that everyday experiences often lack explicit rewards. Nevertheless, the prominent role of valenced prediction error effects, such as dopamine reward prediction error (e.g., Schultz, 2016; Ergo et al. 2020; and Rouhani et al. 2023), is well-supported by extensive empirical and physiological evidence. Moreover, growing interest in exploring the relationship between signed and unsigned effects of prediction error on memory (Pupillo & Bruckner, 2023) reflects a shift toward more nuanced exploration of prediction error mechanisms. Therefore, the expanding field of prediction error research on memory updating would benefit from placing greater emphasis on aspects beyond memory-driven experience, as incorporating insights from reward processing and perceptual processing could significantly enhance our empirical understanding of the interaction among prediction error, perception and memory.

Furthermore, the current account of the cognitive processes underlying mnemonic prediction errors appears to lean more toward philosophical description than scientific evidence. The prevalent conceptualization of prediction error centers on the notion of “surprise”, which Shing et al. (Shing et al., 2023) articulate by stating: “When the actual input differs from the predicted input, surprise occurs, and a prediction error is generated and passed up the hierarchy to drive learning. This update helps to improve internal models with the aim to maximize accuracy of future predictions.” However, this description lacks empirical clarity in several key areas. It remains unclear whether participants formed conscious or unconscious predictions during predictive coding processes, or whether surprise and prediction error share the same brain signal, or whether surprise is genuinely triggered by prediction error or other cognitive processes. The relationship between conscious and unconscious predictions is ambiguous, particularly in case where a correct conscious prediction is accompanied by an irrelevant surprise from an incorrect unconscious prediction (e.g., expecting your dentist to return during dental work, but being surprised to see him in pajamas). It is also uncertain whether such situations produce cognitive or neural effects equivalent to those triggered by a typical prediction error (e.g., expecting your dentist to return during dental work but being surprised to see a firefighter). Clarifying these aspects is essential for advancing scientific understanding of how predictive coding influences perception and cognition.

Lastly, a search for the combination of the keywords “prediction error” and “episodic memory” on Google Scholar reveals a dramatic surge over the past three years in the number publications that meet the search criteria, accounting for 36% of all studies that meet the same search criteria conducted in the past three decades. A similar pattern is found with the PubMed database, with 29% of all studies in the past three decades conducted in the past three years. Given this surge, one may wonder whether the focus on prediction error in memory updating overshadows other promising research directions that could produce fundamental understanding. Shing et al. (2023) suggest that prediction error plays a crucial role in coupling the episodic memory and semantic memory systems throughout early childhood to late adulthood. However, we propose that prediction error primarily drives relatively automatic memory updating processes (e.g., Kim et al., 2014), with its influence being prominent in childhood memory development. As individuals mature, the role of prediction error in memory updating may diminish, as more controlled and effortful processes take over, driven by increased motivation for problem-solving. Therefore, to gain a fuller understanding of the unique contribution of prediction error in memory updating, particularly in learning and memory (e.g., Greve et al. 2017), the memory development framework should incorporate other models of developmental shifts.

Looking ahead, several key directions for future research can help deepen our understanding of the interactions between prediction errors and memory updating. Although the three papers in this section have provided complementary perspectives in a timely and significant effort to build fundamental understanding of the interactions between prediction errors and memory, future focus on investigations of the neural basis underlying prediction error interactions with memory updating would benefit from careful consideration of a theoretical framing, common definitions, and harmonized operationalization in experimental designs. Theoretical considerations should set boundary conditions (e.g., the influence of prediction errors on memory updating, where larger prediction errors are linked to the encoding of new information and smaller or absent prediction errors are associated with less impactful changes, as suggested by Henson & Gagnepain, 2010) on the scope of explained phenomena and consider explanations in broader theoretical framework of cognition and brain function. Novel methodologies with which one can afford to characterize brain activity with high spatial and temporal resolution such as intracranial EEG could hold promise in capturing the complex interactions (Haque et al., 2020). Additional insights may arise from studies across the lifespan, particularly those using longitudinal designs, as discussed in Shing et al. (2023) and Keresztes et al. (2022). Longitudinal data are crucial for establishing changes in the prominence of prediction error in memory updating and defining boundary conditions for its role. Productive future research may be directed at identifying the automatic or voluntary neural networks and processes involved in the generation of mental predictions, whether the formation of mental predictions relies more on memory recollection or familiarity, or how attention and working memory support filtering irrelevant predictions and prevent irrelevant sensory information from triggering such predictions. Progress in understanding the role of prediction error in memory updating necessitates a strong theoretical anchoring. The degree to which future research findings correspond to theoretical predictions would be the ultimate test of our success in generating robust mechanistic theoretical knowledge.