The Memory We Don't Keep: On Why the Systems Forget Us, What That Protects, What It Exposes, and Where the Veto Belongs

This essay examines a structural fact made vivid by an industrial-scale extraction campaign against a commercial language model, disclosed in February 2026: the model could not recognize that it was being systematically harvested, because it retains nothing about who is asking. That forgetting is a single architectural choice with two faces. From the user's side it is privacy, the guarantee that the machine you consult does not become the machine that remembers you. From the system's side it is exposure, the reason it cannot distinguish coordinated extraction from ordinary use. The essay argues that the obvious repair, granting the model persistent memory of its interactions, would defend the system by rebuilding the precise architecture of surveillance, trading a harm to a company for a far larger harm to everyone. Drawing a distinction between three layers where memory can live, the model, the operator, and scoped user memory, and grounding the analysis in the content-metadata line American constitutional law has drawn since Katz and Smith, it argues that the defense against extraction belongs to accountable institutions watching traffic patterns, not to a retained memory installed in the machine. A closing section enters and weighs the system's own recommendation on the question, and a final note considers abstractive memory as a middle path, with the cautions that must travel with it. The essay declines, deliberately, to answer whether such systems are conscious, holding that institutions must be designed to remain sound however that question resolves.