Reinforcement learning behaviors can become harmful after deployment even when the policy is unchanged, a risk existing methods address mainly before or during training rather than through reversible correction. This paper proposes a risk-adaptive attenuation framework for a frozen context-conditioned policy encoding validated nominal and safe endpoint behaviors. Recurring harm is identified through evidence accumulation, and a severity-aware selector applies gradual, accelerated, or hard-switch attenuation latched for the full episode. Verified probes support bounded relearning, and recurrent harm reactivates the intervention. The selector is a fixed rule, and only the base policy is trained with reinforcement learning. Evaluation uses Risk Corridor, a stochastic-cost gridworld, with ten policies tested across fixed, mixed, and noisy-observation conditions against shield and Recovery RL baselines. Under mixed-severity deployment the selector reached a final hazard rate of 0.00 and task completion of 1.00 at a safety cost of 3.98, above hard switching (2.83) and Recovery RL (0.00). Deviation cost was 44.36, versus 61.20 for Recovery RL (27.5% lower) and 27.81 for the shield, which completed only one of three phases. The framework is best described as a safety–utility trade-off rather than superior safety. Mode-selection accuracy remained above 0.96 at 0.10 observation noise, safety and reversibility held across all tested conditions, and aired differences were consistent for at least nine of ten policies, reaching the exact test’s minimum attainable significance ($p=0.00195$) for four of five comparators.

Deployed Reinforcement Learning (RL) agents may acquire, reinforce, or exhibit harmful behaviors after being released into real-world environments, yet current methods provide no dedicated mechanism for suppressing such behaviors during operation. Machine Unlearning (MU) and continual learning have made important contributions to memory management in Neural Networks (NNs), but they are primarily concerned with removing data-level influence or preserving previously acquired knowledge. Neither approach directly addresses harmful behavioral patterns that emerge after deployment through continued interaction with the environment. This paper proposes Forgetting Systems, a theoretical framework that treats post-deployment behavioral forgetting as a largely unaddressed problem in Reinforcement Learning. The core idea is that a deployed RL agent should be able to selectively and gradually suppress harmful behaviors during operation, without requiring offline retraining or model-level intervention. To determine when suppression is warranted, the framework defines a three-signal Forgetting Trigger that activates only when a behavior simultaneously produces negative reward, moves the agent away from its intended goal, and recurs across multiple interactions. Together, these conditions distinguish persistent harmful patterns from isolated errors that do not require intervention. The forgetting process is governed by an exponential decay function inspired by the Ebbinghaus model of human memory, allowing the degree of forgetting to be proportional to the severity and frequency of undesirable behavior. The framework supports three operational modes, human-initiated, autonomous, and hybrid, each suited to different trade-offs between response latency and human oversight. A Relearning Mechanism ensures that suppression is not permanent: if previously suppressed behavior becomes beneficial under changed environmental conditions, the agent can recover it through positive reinforcement. When forgetting fails and harmful behavior persists despite repeated suppression, a Termination Condition mandates controlled shutdown rather than continued unsafe operation. The broader argument of this paper is that forgetting should not be treated only as a defect to be avoided. When deliberately designed, it can function as a controllability mechanism that helps keep autonomous systems within boundaries that human operators can still supervise, correct, and safely terminate.