New ACE Module Boosts LLM Agent Context Management

As AI agents tackle increasingly complex tasks, the length of their operational trajectories grows, posing a significant challenge for large language models (LLMs) with fixed context windows. Traditional context management methods, such as truncation or summarization, are often inflexible and irreversible, meaning crucial information can be permanently lost or compressed.To overcome these limitations, a new module called ACE (Adaptive Context Elasticizer) has been developed. ACE is designed as a plug-and-play component that dynamically manages historical step information within an agent's context. It features a lossless message maintenance layer that stores both raw and compressed versions of past messages, alongside a context orchestration layer that intelligently decides whether to present each historical step as raw, abstract, or to drop it, based on the current task state.This reversible design ensures that the LLM always receives a compact yet information-rich context. ACE was successfully adapted to four different agent frameworks—ReAct, DeepAgent, WebThinker, and MiroFlow—without requiring any training or architectural modifications. Experiments demonstrated that ACE consistently outperformed existing truncation and summarization baselines, delivering performance gains across all tested frameworks.