Large language models (LLMs) struggle to incorporate new knowledge without either forgetting previously learned information or requiring expensive, time-consuming retraining. This limitation is particularly acute in dynamic domains like supply chain management, where regulations, supplier data, and trade policies change frequently. A new research paper introduces DYNA (Dynamic Episodic Memory Networks), a framework designed to address this challenge by augmenting a frozen LLM with an external, updatable memory source.
According to the arXiv paper by Sarabadani and Tajvidiyan, DYNA uses a temporal knowledge graph where events are represented as nodes and temporal relations are directed, timestamped edges. This graph acts as an episodic memory that can be updated without modifying the underlying LLM. At query time, DYNA retrieves relevant nodes using random walks and centrality measures, then feeds that context into the LLM to generate a response.
How DYNA Works
DYNA keeps the LLM frozen, avoiding the computational cost of fine-tuning. Its core innovation is the temporal knowledge graph, which stores facts as time-stamped events. The retrieval mechanism relies on graph traversal and centrality scoring to pull relevant subgraphs. This approach allows the system to answer questions that require temporal ordering or recent information without catastrophic forgetting.
Performance Gains
The authors evaluated DYNA on three temporal recall tasks and compared it against fine-tuning and standard retrieval-augmented generation (RAG). The results are summarised below.
| Metric | DYNA vs Fine-Tuning | DYNA vs Standard RAG |
|---|---|---|
| Catastrophic forgetting reduction | ~7% less forgetting | — |
| Temporal ordering accuracy improvement | — | ~5% better ordering |
The paper also reports that higher graph clustering coefficients correlate with better retrieval performance, suggesting that the structure of the knowledge graph itself matters for effective retrieval.
Implications for Enterprise AI
For enterprise technology buyers, the key takeaway is that DYNA demonstrates a retraining-free path to keeping LLMs current. In supply chain and trade contexts—where regulations, supplier statuses, and tariff rates shift regularly—a frozen model that can incorporate new data through an updatable graph could significantly reduce the cost and latency of model updates. While the paper is academic and does not mention specific enterprise applications, the principle of an external temporal memory is relevant to any use case requiring continuous learning from time-sensitive data.
DYNA's reliance on graph properties also implies that investing in well-structured knowledge graphs could improve the performance of augmented LLMs. The framework itself is lightweight, meaning it could feasibly be integrated into existing LLM deployments without major hardware or software overhauls.
As LLMs become more prevalent in logistics, customs, and trade finance, solutions like DYNA could help bridge the gap between static training and dynamic reality—without the disruption of constant retraining.