DYNA : Dynamic Episodic Memory Networks for Augmenting Large Language Models with Temporal Knowledge Graphs in Continuous Learning

DYNA is a proposed framework designed to enhance frozen large language models (LLMs) by integrating a temporal knowledge graph that facilitates continuous learning without retraining. It employs random walks and centrality measures for retrieving relevant nodes from the graph, resulting in a ~7% reduction in catastrophic forgetting and a ~5% improvement in temporal ordering on three evaluated tasks. This approach highlights the significance of graph structure in retrieval performance, making it a valuable tool for practitioners aiming to maintain LLMs' knowledge over time.