The institutional problem is that AI companies have made memory sound like a user-controlled feature while selling assistants that feel valuable because they carry context forward. That bargain also appears in applied-agent case studies, where operators stress boundaries, observability, and escalation because misclassified calls become real problems outside the demo. ChatGPT's memory complaint puts the same pressure on consumer AI: the control panel is now part of the trust claim, and users are testing it through ordinary conversation.