ChatGPT's new memory upgrade is powerful - and could poison every answer it gives you
Original reporting by ZDNet
OpenAI has recently rolled out what it calls significant "improvements" to ChatGPT's memory, a feature designed to create a more personalized user experience by remembering past conversations. But while OpenAI celebrates its advanced "Dreaming V3" capability – which autonomously builds comprehensive user profiles from chat histories and inferred preferences – one ZDNET editor warns that these changes are deeply unsettling, raising more concerns than they resolve.
Previously, ChatGPT's memory was a basic list of facts, often quickly becoming irrelevant. The new system, now available to Plus/Pro subscribers and rolling out to all users, actively synthesizes every interaction to develop a detailed "dossier" on the user. OpenAI hails this as a breakthrough in efficiency and recall, with metrics showing a substantial jump in factual task success and preference adherence.
Unpacking the concern Yet, the author's testing reveals a troubling reality. ChatGPT's advanced memory frequently misinterprets details, clinging to outdated information and even fabricating false connections, such as inaccurately claiming the author uses Home Assistant. This automated profiling, while technically impressive, leads to responses filtered through an often-flawed internal representation of the user. Beyond inaccuracy, concerns mount over privacy and user control, as it's nearly impossible to prune what the AI decides to remember or infer. Even disabling the memory feature may not fully erase the AI's gathered insights, leaving users with unsettling questions about transparency, data retention, and the potential for AI to skew information based on an imperfect, uneditable self-perception.
The concerns raised about ChatGPT’s evolving memory capabilities represent more than mere technical glitches; they mark a fundamental, and potentially problematic, shift in the human-AI interaction paradigm. While OpenAI champions advancements in factual recall and operational efficiency, the practical experience for users reveals a system grappling with accuracy, frequently misinterpreting personal context, and providing inadequate mechanisms for genuine data control. The notion of each chat as a clean slate is rapidly eroding, replaced by an always-on profiling mechanism that, despite its sophisticated underpinnings, struggles to truly understand its human counterpart. This creates a significant cognitive burden, forcing users to constantly scrutinize responses against an unseen, often flawed, internal dossier the AI has constructed, thus inverting the very promise of AI to reduce human effort.