The paper presents a control problem in persistent LLM memory, proposing a separation between utility and authority to address the issue. It introduces rank-normalized compensation as a scale-invariant solution and evaluates its effectiveness on the Memora Remembering dataset.