LLMs are not good strategists, but a memory of successful episodes changes the game

The EpicStar framework combines working memory with a bank of winning episodes; on StarCraft II, it wins more often by consuming an order of magnitude less tokens.

On long-term tasks, language models derive: their attention is over, and the coherence of a plan on hundreds of decisions ends up breaking down. Yi Wu and Zhimin Hu address this problem in an article on arXiv on August 12, 2026.

Their proposal, called EpicStar, keeps in a double memory register. In addition to the usual work memory is a bank of successful past episodes, from which the agent can draw. A dynamic selection mechanism slices each step: replaying an action already proven in a similar situation, or initiating a new reasoning.

Evaluated on StarCraft II, a demanding terrain for long-term planning, EpicStar surpasses the reference methods in victory rates while consuming an order of magnitude of less tokens — and retains this advantage at different levels of difficulty as compared to various opposing strategies. The result suggests that the observed strategic weakness is due less to a lack of reasoning than to a retention problem.