Knowledge graph - Scientific news
A research team at MIT has developed "Graph--State Memory (GSM)," a new architectural component for Large Language Models that replaces traditional KV--caches with a dynamic knowledge graph. Presented on May 23, 2026, this breakthrough allows LLMs to maintain long--term factual consistency across millions of tokens by storing information as structured relationships rather than raw embeddings. For the business sector, GSM offers a solution to the "context window decay" problem, enabling AI agents to manage multi--year project lifecycles or complex legal cases without losing track of critical historical dependencies.