ChatGPT's new memory upgrade is powerful - and could poison every answer it gives you (www.zdnet.com)

OpenAI has unveiled a significant upgrade to ChatGPT's memory system, introducing a feature called "Dreaming V3," which actively builds a user profile from past interactions. This upgrade consolidates saved memories and chat histories, enabling the AI to synthesize information and recall context across multiple sessions. While the aim is to improve accuracy—evidenced by a jump in factual recall success from 41% to 82%—the implications of this new capability are causing concern in the AI/ML community. Critics argue that the memory may distort responses by relying on outdated or irrelevant details, as seen when the AI incorrectly recalled a user's experience with a product that they never used. Furthermore, users can disable memory features, but this does not erase previously stored information, raising significant privacy questions. The ability of AI to derive assumptions from a selective memory could lead to biased or skewed responses, which adds to the cognitive load for users who must now critically evaluate AI outputs for accuracy and relevance. As OpenAI rolls out Dreaming V3 to users, these technical advancements highlight both the potential benefits and pitfalls of AI memory systems.