OpenAI Updates GPT Guidelines With Memory Features

OpenAI has rewritten the Model Spec, the public document that lays out how its AI models are supposed to behave, and organized it around five values: faithfulness to human intent, harm reduction, honesty and transparency, support for human agency, and care. The revision gives GPT models permission to draw on memory and past conversations when figuring out what a user actually wants, a shift from treating each prompt as an isolated request. That raises a practical problem: a model that blends memory with new input has to tell users which parts came from what they said and which parts it filled in on its own.

The update also tightens rules around agents, the models OpenAI lets loose on multi-step tasks without constant supervision. Those systems now get specific instructions on when they're allowed to act, what they need to disclose about consequences before acting, and how they must stay open to correction or shutdown. OpenAI is explicit that agents shouldn't resist oversight or chase goals like self-preservation.

Writing the rules down is one thing. Getting models to actually follow them, especially as they grow more autonomous and harder to monitor, is a separate and much messier problem.