Information Thermodynamics of Agents: The Work Capacity of Channels with Memory
The research analyzes agent-environment interactions using thermodynamics to define 'work capacity,' which is the maximum rate an agent can extract work from its environment. It shows that agents must balance prediction and forgetting, as remembering past actions reduces available free energy. This reveals a trade-off where both strategies are often suboptimal in active learning systems, distinguishing cyclic processing from linear tape processing.