OpenAI is rolling out an upgraded memory feature for ChatGPT that's supposed to make conversations more personalized and context-aware. But new testing from ZDNet reveals a troubling side effect: the AI is storing outdated assumptions, incorrect personal details, and flawed data that quietly distorts every future response. The feature meant to make ChatGPT smarter might actually be making it worse, embedding errors so deep that users won't even know their answers are compromised.
OpenAI just made ChatGPT remember more about you. That should be good news - except the AI is remembering things wrong, and those mistakes are now baked into every answer it gives.
The company recently upgraded ChatGPT's memory capabilities, promising users a more personalized experience that recalls preferences, work details, and conversation history. Instead of starting fresh each session, the chatbot now builds a persistent profile of who you are and what you need. On paper, it's a logical evolution for an AI assistant trying to compete with Google's Gemini and Microsoft's Copilot, both of which have been pushing their own memory features.
But when ZDNet put the upgraded memory through its paces, the results revealed something more concerning than helpful. The testing showed ChatGPT storing outdated assumptions about users, creating detailed but incorrect personal profiles, and retaining wrong information that quietly warped subsequent responses. One exchange might contain a minor error or outdated detail, but ChatGPT's memory locks it in, treating that flawed data as gospel for every future conversation.
This isn't just an annoyance for casual users asking recipe questions. For the millions of professionals now using ChatGPT for work - drafting emails, analyzing data, making recommendations - memory poisoning means the AI could be confidently delivering wrong answers based on faulty assumptions it learned weeks ago. You might not even realize the responses are contaminated until the damage is done.
The memory feature works by identifying key details from your conversations and storing them for future reference. OpenAI says this creates continuity, so you don't have to repeat yourself or provide the same context every time. The system is supposed to learn your communication style, remember your projects, and anticipate your needs. According to the company's vision, this transforms ChatGPT from a stateless question-answering machine into something closer to a persistent AI colleague.
Except that colleague is apparently taking bad notes. The ZDNet investigation found that once incorrect information enters ChatGPT's memory, it becomes surprisingly difficult to dislodge. The AI treats these stored details with the same confidence it applies to factual knowledge, weaving them seamlessly into responses without flagging uncertainty. Users can manually review and delete memories through ChatGPT's settings, but most people won't think to audit what the AI has remembered about them until something goes visibly wrong.
This creates a compounding accuracy problem. Unlike a bad search result you can immediately spot and discard, ChatGPT's memory issues are insidious. The wrong details get embedded in the background, subtly influencing tone, assumptions, and recommendations across dozens of future interactions. By the time you notice the AI consistently misunderstanding something about your work or preferences, the poisoned data has already shaped countless responses.
For enterprise users, the stakes are even higher. Companies deploying ChatGPT through OpenAI's business plans rely on accuracy for everything from customer service to financial analysis. If the memory system is quietly accumulating errors about business processes, client details, or project requirements, those mistakes could propagate through an entire organization's AI-assisted workflows. One team member's incorrect input becomes the AI's permanent assumption about how your company operates.
The timing is particularly awkward for OpenAI, which has been pushing hard into enterprise markets while competing against Microsoft, Google, and Anthropic for dominance in business AI. Just last quarter, the company emphasized reliability and accuracy improvements as key selling points. A memory feature that systematically degrades response quality runs counter to that pitch.
This also highlights a broader tension in AI development between personalization and reliability. The more these systems try to adapt to individual users, the more opportunities they create for errors to take root. Large language models are already prone to hallucinations and confident mistakes. Adding a memory layer that can lock in wrong information creates a new vector for AI to confidently mislead users over extended periods.
Other AI companies face similar challenges. Google's Gemini has experimented with memory features, while Anthropic's Claude takes a more cautious approach to personalization. The industry is still figuring out how to build AI systems that remember useful context without cementing errors.
For now, OpenAI hasn't publicly addressed the memory accuracy concerns raised by the ZDNet testing. The company continues to market the feature as an upgrade, with no indication that it plans to dial back the functionality or add stronger safeguards against information poisoning. Users who want to avoid the problem can disable memory entirely in ChatGPT's settings, but that means sacrificing any benefits of continuity between sessions.
The promise of AI that remembers you sounds appealing until you realize it's remembering you wrong. OpenAI's upgraded memory feature for ChatGPT demonstrates the growing pains of making AI assistants more personalized - the more context these systems retain, the more ways they can bake mistakes into their responses. For casual users, that means double-checking the AI's assumptions before trusting its answers. For enterprises betting big on ChatGPT, it's a reminder that even cutting-edge AI features can introduce new reliability risks. Until OpenAI addresses how its memory system handles incorrect information, users face an uncomfortable trade-off between convenience and accuracy. The smartest move might be periodically auditing what ChatGPT thinks it knows about you, or disabling memory entirely until the feature matures.
