Long-horizon tasks require AI agents to interact with tools over many steps, but the growing history of observations and failed attempts quickly overloads the agent’s working memory, degrading its ability to plan and reason.
The COgnitive Resource Self-ALlocation (CORAL) framework gives agents a set of tools to manage their own memory. The agent can checkpoint its progress and verified facts, then reset its context to a clean state, preserving only the checkpointed information. This way it starts each episode fresh.
Tested on the GAIA benchmark, CORAL achieved 42.8% average accuracy on the hardest tasks, using six to seven times fewer input tokens than a standard method. Better memory management made the agent both more accurate and more efficient.