Large language model agents can solve complex questions, yet they struggle when a task requires many steps. The problem is cognitive overload: the agent’s working memory fills with irrelevant data and its attention spreads. The authors introduce Cognitive Resource Self‑Allocation (CORAL) to let agents manage their own memory.

CORAL gives agents two tools they can call during reasoning. The Memory Management tool lets an agent take a checkpoint, saving its progress and verified facts. The Context Optimization tool can then erase useless information from its memory while keeping the checkpoint. The agent starts a new problem‑solving episode from a clean state, resuming once the useful facts are restored.

On the GAIA benchmark, which tests long‑horizon reasoning, an 8‑B CORAL model scores 42.8 % average accuracy. It outperforms a ReAct agent and uses 6–7 times fewer input tokens. CORAL also improved performance on multi‑hop knowledge‑QA datasets and on the harder BrowseComp benchmark, where it raised a strong proprietary agent’s score by 8.0 points.

These results show that letting a language‑model agent control its own memory can reduce overloading and save computation, especially for tasks that need many reasoning steps. The gains come without changing the underlying model architecture and keep inference costs lower than standard methods.