In some environments, an agent cannot both retain every detail needed for optimal prediction and achieve the highest possible energy efficiency. A 2026 theoretical study finds that discarding parts of an agent’s own action history can improve the useful energy it extracts through repeated interactions—even when the discarded information would help predict what happens next. The result is a mathematical finding about certain models, not evidence that today’s AI systems have this limitation.
What the study found
The study examines an agent that observes an environment, acts on it, and then observes the consequences. Unlike a passive information processor, this agent changes the world whose future it is trying to predict. Its actions become part of the history that can inform later predictions.
| # | Preview | Product | Price | |
|---|---|---|---|---|
| 1 |
|
Basic Physics: A Self-Teaching Guide, 3rd Edition | $10.48 | Buy on Amazon |
| 2 |
|
Fundamentals of Physics I: Mechanics, Relativity, and Thermodynamics (Open Yale Courses) | $35.00 | Buy on Amazon |
| 3 |
|
Must Know High School Physics | $14.57 | Buy on Amazon |
| 4 |
|
Physics | $41.26 | Buy on Amazon |
| 5 |
|
Physics for Scientists and Engineers With Modern Physics | $91.26 | Buy on Amazon |
As an Amazon Associate I earn from qualifying purchases.
For some environments, the authors prove that an agent retaining all information relevant to optimal prediction cannot attain the maximum possible energy efficiency. The objective is the greatest average useful energy extracted per interaction over repeated exchanges with the environment. In those cases, retaining prediction-helpful details of past actions can conflict with maximizing that thermodynamic objective.
The coverage of the study appeared in Phys.org on October 9, 2026, with research coverage credited to the University of Innsbruck. The paper is Lukas J. Fiderer et al., “Information Thermodynamics of Agents: The Work Capacity of Channels with Memory,” published in Physical Review X in 2026 (DOI: 10.1103/7nds-tjr8). The result should be read as a mathematical claim for the environments covered by the work, not a rule for every intelligent system.
#1 Best Overall
Why an agent’s actions matter
When an agent only receives information, its observations do not depend on actions it took. An acting agent faces a feedback loop: it uses its history to choose an action, the action affects the environment, and the resulting observation becomes new evidence. The past can therefore be useful both for predicting the environment and for determining how the agent interacts with it.
The study extends earlier work on passive information processing to this setting in which actions shape later observations. As lead author Lukas Fiderer put it in the coverage, “Through his actions, the agent changes the world he himself is trying to predict. We wanted to understand what that changes about the physics of information processing.”
What “forgetting” means in this result
Here, forgetting means not retaining some details of the agent’s own action history that could help predict future observations. It does not mean the agent runs out of storage or loses information because a practical memory system is too small. The reported effect persists with additional memory capacity and is independent of practical memory-storage costs.
The Tool Desk
Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →The coverage illustrates the idea with a robot opening a door. Remembering how hard it pushed could help the robot predict how the door will move. In the model’s thermodynamic framing, however, forgetting that force can be beneficial in some environments. This is an illustration, not a reported robot experiment.
Rank #3
Fiderer described the counterintuitive point this way: “The surprising part is that an agent can benefit energetically from forgetting information that would actually improve its predictions.” The claim concerns the energy objective in the model; it does not say that forgetting generally improves prediction or that an agent should discard useful data in ordinary applications.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What the result does—and does not—say about AI
The work identifies a possible fundamental limit for ideal agents in certain environments. It does not report measured energy savings, a drop in prediction accuracy on a deployed AI system, or a test on a physical robot. The coverage gives no numerical estimate for energy or predictive performance.
Rank #4
It also does not establish that the finding can be translated directly into lower energy use or more sustainable AI. Fiderer is quoted as saying that current computers remain far from these fundamental limits and that the work offers a way to explore what efficient agents could look like as technology advances. The coverage presents a proof-of-principle experiment as a possible future step, not as a result already demonstrated.
For the theorem’s assumptions and exact conditions, consult the published paper; the news coverage alone does not provide enough detail to reconstruct them. The practical significance is therefore conceptual for now: when an agent’s actions alter what it later observes, the best memory for prediction need not be the memory that optimizes the model’s energetic objective.
Quick Recap
Best Value
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
