Predictions transform memories: How expected versus unexpected events are integrated or separated in memory

Abstract

Our brains constantly generate predictions about the environment based on prior knowledge. Many of the events we experience are consistent with these predictions, while others might be inconsistent with prior knowledge and thus violate our predictions. To guide future behavior, the memory system must be able to strengthen, transform, or add to existing knowledge based on the accuracy of our predictions. We synthesize recent evidence suggesting that when an event is consistent with our predictions, it leads to neural integration between related memories, which is associated with enhanced associative memory, as well as memory biases. Prediction errors, in turn, can promote both neural integration and separation, and lead to multiple mnemonic outcomes. We review these findings and how they interact with factors such as memory reactivation, prediction error strength, and task goals, to offer insight into what determines memory for events that violate our predictions. In doing so, this review brings together recent neural and behavioral research to advance our understanding of how predictions shape memory, and why.

Keywords: Integration; Learning; Memory; Prediction; Prediction error; Prior knowledge; Schema; Separation.

Figure 1.

Prior knowledge allows us to make memory-based predictions about novel events, such as seeing a toaster and expecting that toast will soon follow. These predictions can either be met (i.e., consistent with prior knowledge) or violated (i.e., inconsistent with prior knowledge, such as seeing a frog in your kitchen next to the toaster).

Figure 2.

Hypothesized neural and mnemonic outcomes for events that meet (a) vs. violate (b) our predictions. We conceptualize representations of elements of experiences as nodes with links between them, illustrated here by circles with connecting lines, respectively. These nodes and links can be thought of as belonging to a theoretical neural network or an associative network model, or as neurons (nodes) and synapses (links) in neural ensembles. The top illustrations reflect representations active during encoding. Representational changes can occur already during encoding, or during consolidation or retrieval. (a) When an experience conforms to our predictions, neural integration — that is, increased overlap between memory representations — is likely to occur (resulting in neural similarity). We hypothesize that increased overlap occurs via the strengthening of existing links between nodes representing elements of an experience, or via the creation of new links, potentially leading to more shared nodes. Integration can be either between different elements of the experience (e.g., between novel images of a toaster and toast), or between the novel experience and existing knowledge structures (e.g., between the novel association and your general knowledge about kitchens). Integration can promote both enhanced associative memory as well as increased susceptibility to false memories for other knowledge-consistent details. (b) In contrast, when events violate our predictions, both neural integration and separation — whereby memory representations become less similar/overlapping — can occur. Potential factors influencing separation or integration in response to prediction errors are illustrated in the black inset. Separation can happen via multiple mechanisms, such as creating a new representation of the novel experience (toaster-frog association), fully distinct from that of the pre-existing memory prediction (toaster-toast association), or by weakening links between nodes or inactivating nodes initially shared by both elements, resulting in fewer shared nodes (not illustrated here). Integration may promote the updating of prior knowledge (to accommodate the previously unexpected information). Separation may enhance memory for the violating information itself, and lead to distinct memories for both the prior prediction and the unexpected experience. Question marks reflect that these hypotheses remain to be empirically tested. Potential factors that might shape the fate of memories for knowledge-inconsistent information towards either integration or separation are illustrated in the black inset box (see main text for details).

Figure 3.

Illustrative examples of paradigms used to test memory for information that is consistent versus inconsistent with predictions or prior knowledge. (a) Information is either congruent or incongruent with existing semantic knowledge, and therefore is either consistent or inconsistent with predictions arising from semantics. (b) In an AB/BC learning paradigm, participants first learn an association between two items, A-B, and then subsequently see one of those items, B, presented with a new item, C, thereby violating their predictions of having item B associated with item A. (c) In one example of novel information being added to a structured schema, participants first learn a spatial schema (a grid of object locations), then encode grids of novel objects that are either consistent with that spatial schema (learned in the same locations) or inconsistent with it (different locations). This kind of paradigm has also been done using non-spatial schemas, as well as pre-existing schemas learned through lifetime exposure. (d) In a sequential learning task, participants first learn sequences of stimuli through repeated exposure, and are then presented with sequences that either conform to or violate the learned sequence. Note that some previous studies combined elements across paradigms, e.g., sequential learning tasks in which item sequences are consistent or inconsistent with semantic knowledge (e.g., “theater – popcorn – candy” vs. “theater – popcorn – basketball”).