Pruning of memories by context-based prediction error

The capacity of long-term memory is thought to be virtually unlimited. However, our memory bank may need to be pruned regularly to ensure that the information most important for behavior can be stored and accessed efficiently. Using functional magnetic resonance imaging of the human brain, we report the discovery of a context-based mechanism for determining which memories to prune. Specifically, when a previously experienced context is reencountered, the brain automatically generates predictions about which items should appear in that context. If an item fails to appear when strongly expected, its representation in memory is weakened, and it is more likely to be forgotten. We find robust support for this mechanism using multivariate pattern classification and pattern similarity analyses. The results are explained by a model in which context-based predictions activate item representations just enough for them to be weakened during a misprediction. These findings reveal an ongoing and adaptive process for pruning unreliable memories.