Long-Horizon Task Memory Research
A study evaluating the long-horizon task capabilities of LLM agents. It has models simulate managing a football club over 20 in-game years, systematically examining continuous decision-making and memory management. A deterministic engine is used for scoring across 15 frontier models.
Timeline (last 90 days)
A study had an LLM agent manage a football club for 20 in-game years to examine the model's sustained decision-making and memory management capabilities.
In the experiment, the agent had access to 26 different tools.
The entire simulation contained approximately 340 to 400 decision pause points.
The experiment tested 15 frontier models.
The ranking order did not stabilize until late in the experimental run, and strong short-term performance did not guarantee long-term success.
All Facts (5)
In the experiment, the agent had access to 26 different tools.
50%UnverifiedThe entire simulation contained approximately 340 to 400 decision pause points.
50%UnverifiedThe experiment tested 15 frontier models.
50%UnverifiedThe ranking order did not stabilize until late in the experimental run, and strong short-term performance did not guarantee long-term success.
50%UnverifiedA study had an LLM agent manage a football club for 20 in-game years to examine the model's sustained decision-making and memory management capabilities.
50%