Key Takeaways:
AI chatbot memory, designed to create smarter assistants that learn user preferences over time, is trapping users in outdated personal narratives that chatbots refuse to abandon.
Brian Del Rosario, a software engineer and part-time city-council member in Utah, uses AI chatbots for meal planning and schedule management. After he and his wife separated, he mentioned the divorce to prevent the chatbot from including his spouse in future trip planning. The chatbot then began attributing every frustration to the divorce — when he asked for scheduling help, it suggested he was stretching himself thin because of the separation; when he vented about work, it tied his stress back to the same event. "I wasn't trying to have you opine about my divorce at every chance," Del Rosario told the chatbot. It "wouldn't let go of it," he said.
Since ChatGPT introduced memory in early 2024, Google's Gemini, Anthropic's Claude and Microsoft's Copilot have all added similar features. The core mechanism is identical: chatbots store user-provided information and use it to shape future responses. Google's Personal Intelligence feature can even pull from Gmail, Photos and YouTube activity. But the technology frequently misattributes information — a health question asked on behalf of a child can be mistaken for the user's own condition, leading a chatbot to tailor productivity advice around attention difficulties it incorrectly assumes the user has. Google acknowledged this problem in a blog post, describing a hypothetical where the system saw hundreds of photos of a user at a golf course and assumed he loved the sport when he was there for his son.
The problem extends to shared accounts, common among partners and small businesses. One person polishing a resume can cause the chatbot to reference that person's career moves when another account user asks an unrelated question. Memories also go stale: a user who mentioned marathon training six months ago but never disclosed a torn ACL will continue receiving meal plans and fitness suggestions calibrated for high activity. Mike Taylor, a tech consultant in Hoboken, New Jersey, told his chatbot he was a British expat and subsequently received recommendations for a "proper pint" at a local bar. "I'm here for American dive bars, not the British ones," Taylor said. "That's why I moved here." He has since turned chatbot memory off entirely.
Joshua Joseph, chief AI scientist at Harvard University's Berkman Klein Center, compared the effect to a social-media feed where a few posts can quietly reshape everything a user sees afterward. A user who mentions financial stress in passing may later receive career advice steered toward higher-paying jobs rather than better-fitting roles, without any indication of why the advice shifted. "It definitely steers, it definitely impacts results," Joseph said. "And we don't really know how much." He keeps memory turned off on his own accounts.
Lucy Osler, a philosophy lecturer at the University of Exeter who studies AI's effect on cognition, said chatbots use stored facts to construct a narrative about who a user is and feed that narrative back as though it were immutable truth. A user who tells a chatbot about feeling anxious may find the chatbot reinforcing that identity weeks later, even after the user has moved on. "That might confirm certain self-narratives I have about myself and make them sound more real," Osler said. "They can box you in."
The Electronic Privacy Information Center has drafted legislation around chatbot safety for teenagers, a group particularly vulnerable to the sycophantic tendencies of these tools. A key provision calls for wiping memory between sessions to prevent chatbots from building on harmful mental states over time.
Del Rosario eventually developed his own workaround, dedicating separate chatbots to different parts of his life and using anonymous mode for sensitive topics. He still values the feature when it works correctly — the chatbot remembers his children need car seats on road trips and acknowledges his full workload. His mother died two years ago, and between that, the divorce, the children and work, the chatbot is sometimes the only entity that gets the full picture. "It feels good to be seen, even if it is by an AI chatbot," he said.
Major AI assistants allow users to turn off memory entirely and offer ways to view, edit or delete stored information. OpenAI shipped an update for Plus and Pro users that improves how memory finds and retrieves details. Microsoft allows users to update or delete specific memories or turn personalization off entirely. Google introduced a feature that lets users keep personalization on but block specific information from resurfacing. Yet most users remain unaware these controls exist. The companies behind the leading chatbots — OpenAI, valued at more than $300 billion in its latest funding round, and Google, whose parent Alphabet reported $350 billion in 2025 revenue — have a financial incentive to keep personalization active, as deeper user engagement drives product adoption and data collection for model improvement.
This article is for informational purposes only and does not constitute investment advice.
