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

A new framework called DYNA pairs a frozen large language model with a temporal knowledge graph to let the model absorb new information without retraining, according to a paper posted to arXiv on 14 June 2026 [1]. Large language models typically struggle to incorporate new knowledge without forgetting earlier material or undergoing expensive retraining cycles [1][2]. The authors propose DYNA, which treats events as nodes and temporal relations as directed, timestamped edges inside an external, updatable memory [2]. At query time, the system retrieves relevant nodes through random walks and centrality measures, then augments the frozen model’s response [2]. On three temporal recall tasks, DYNA cut catastrophic forgetting by roughly 7 percent compared with fine-tuning and improved temporal ordering by about 5 percent over standard retrieval-augmented generation [2]. The paper also reports that higher graph clustering coefficients correlate with better retrieval performance, suggesting that the structure of the knowledge graph directly influences accuracy [2]. The work lands as the broader AI community continues to search for methods that keep models current without full retraining. Large language models are defined as machine-learning systems with many parameters, trained through self-supervised learning on vast text corpora [8]. Recent entrants such as DeepSeek and Qwen have drawn attention for lowering training costs and releasing open-weight variants, yet the fundamental problem of knowledge staleness persists across architectures [7][9]. DYNA’s approach of using a temporal knowledge graph as episodic memory is retraining-free, meaning the base model’s weights stay frozen while the graph absorbs new facts [2]. The paper lists three contributions: framing episodic memory as a temporal knowledge graph, enabling LLM augmentation without retraining, and identifying graph properties as predictors of retrieval performance [2]. The preprint appears on arXiv, a repository that accounts for roughly 95 percent of paper URLs organically linked by Hugging Face users [4]. Hugging Face’s paper-pages feature lets researchers link models, datasets, and interactive demos to arXiv preprints, and the platform has collaborated with arXiv to embed demos directly on abstract pages [4][5]. The DYNA paper does not yet list associated models or demos on the Hub, though the integration pathway is well documented [5].