AdaCoM (Adaptive Context Management) trains an external LLM to manage context for a "frozen" agent. The system utilizes reinforcement learning (RL) to prune stale content and preserve task constraints, reducing the degradation of reasoning capabilities in long-context scenarios.
Why It Matters
It enables existing agents (even closed-source ones) to handle deeper research or more complex web searches without requiring retraining.
