Non-parametric memories are external data stores, such as text corpora, document collections, or knowledge bases, that machine learning models access at inference time through retrieval mechanisms to obtain factual information. Unlike parametric memory, which stores knowledge implicitly within a neural network's trained weights, non-parametric memory exists outside the core model architecture. This separation enables models to dynamically query, expand, and update their accessible knowledge base without requiring parameter fine-tuning or retraining, facilitating access to rare, changing, or domain-specific facts while reducing the burden of storing all world knowledge directly in network parameters.
