Brian Del Rosario, a software engineer and part-time city-council member in a small town in Utah, uses AI chatbots for everything from meal planning to managing his schedule. In some of those conversations, he revealed he had a spouse and three children.
Then, after he and his wife separated, Del Rosario had to mention it to the chatbot so it wouldn’t include his wife when planning a future trip. But once he did, the chatbot latched onto the divorce.
When he asked for help managing his schedule, it suggested he might be stretching himself thin because of the divorce. When he vented about a frustrating day at work, it tied his stress back to the divorce.
He says he told the chatbot, “I wasn’t trying to have you opine about my divorce at every chance.” The chatbot, says Del Rosario, “wouldn’t let go of it.”
One of the best things about chatbots is that they have a long memory, learning more about you from one conversation to the next. The result is a smarter assistant that knows your writing style, remembers your dietary restrictions or picks up a project where you left off.
But that great memory also carries some drawbacks. It can get stuck on misunderstood or outdated information. It can feel invasive. And it can make it harder to turn a new page in your life.
In other words, you may be over it, but your chatbot isn’t.
Forever facts
Since ChatGPT introduced memory in early 2024, Google’s Gemini, Anthropic’s Claude and Microsoft’s Copilot have all added their own versions. Their approaches differ, but the basic idea is the same: The chatbot remembers what you have told it and uses that to shape future responses. Google’s Personal Intelligence feature can even pull from a user’s Gmail, Photos and YouTube activity.
One problem: It may be remembering things that aren’t about you at all. For instance, a health question asked on behalf of a child or parent can be mistaken for your own. Ask about symptoms of ADHD for your child, and weeks later when you ask for productivity advice, the chatbot may tailor its suggestions around attention difficulties it thinks you have.
Google, an Alphabet unit, acknowledged a version of this problem in a blog post about its Personal Intelligence feature, describing a hypothetical case where the system saw hundreds of photos of a user at a golf course and assumed he loved the sport, when really he was just there for his son. (In the hypothetical, the user had to tell the chatbot he doesn’t like golf.)
Google says it has introduced a feature that lets users keep personalization on but block specific information from resurfacing in conversations with the chatbot. OpenAI, meanwhile, says it has shipped an update for Plus and Pro users that improves how memory finds and retrieves details. Microsoft says that users can update or delete specific memories (or everything, if they choose). They can also turn personalization and memory off entirely at any time. Anthropic declined to comment.
In a shared account, common enough between partners or in small businesses, the risks multiply. One person uses the chatbot to polish a résumé. Later, when someone else on the account asks an unrelated question, the chatbot might reference that person’s recent career move or suggest next steps for a job search as part of the answer.
Memories can also go stale. Say you told the chatbot you were training for a marathon six months ago. Since then you tore your ACL, but you never mentioned that. Now every meal plan and fitness suggestion is calibrated for someone who’s highly active. You’re following advice built for a version of you that no longer exists.
A real downer
Del Rosario experienced something similar. He had mentioned he was trying to lose weight, and the chatbot started bringing up that fact everywhere, including when he was looking for restaurant recommendations while out of town.
“It was like, ‘Thanks for being the buzzkill about my vacation,’ ” he says. “I wasn’t actually planning to stick to my diet on this trip.”
Similarly, Mike Taylor, a tech consultant for media and software company Every in Hoboken, N.J., once mentioned to a chatbot that he was a British expat. Subsequently, he says, the chatbot recommended a “proper pint” at a local bar, a tip he didn’t find useful. “I’m here for American dive bars, not the British ones,” he says. “That’s why I moved here.”
Taylor has turned his AI chatbot memory off, so that he knows exactly what’s shaping how it responds to him. “The LLM [large language model] doesn’t know who you are, and therefore it won’t bias the results you get,” he says.
Indeed, AI memory can shape results in ways that can be hard to detect. Joshua Joseph, the chief AI scientist at Harvard University’s Berkman Klein Center, compared the effect to a social-media feed and the way a few posts you linger on can quietly reshape everything you see afterward.
Say you mention you’re stressed about money in a conversation about something else entirely. Weeks later, you ask the chatbot for career advice, and it steers you toward higher-paying jobs rather than roles that might be a better fit, because it “knows” you’re financially anxious. You would never know why the advice felt off, because the chatbot hasn’t flagged which memories it is drawing on.
“It definitely steers, it definitely impacts results,” Joseph says. “And we don’t really know how much.” He keeps memory turned off on his own chatbot accounts.
A chatbot that remembers everything can also make it harder for people to move on from their own past. Lucy Osler, a philosophy lecturer at the University of Exeter who studies how artificial intelligence shapes cognition, says chatbots use facts to construct a narrative about who you are, and feed that narrative back to you as though it were fact. If you tell your chatbot you’re feeling insecure and anxious, that becomes how it sees you, and may keep reminding you of that even if you’ve moved on.
“That might confirm certain self-narratives I have about myself and make them sound more real,” Osler says. “They can box you in.”
Negative patterns
Being reminded you were anxious months after the fact can be upsetting. But it can also do real harm. Chatbots are designed to be agreeable, to build on your version of reality rather than question it. Osler says this makes chatbots capable of reinforcing delusional thinking.
This concern has led the Electronic Privacy Information Center to draft legislation around chatbot safety for teenagers, a population particularly vulnerable to the sycophantic tendencies of these tools. A key provision calls for wiping memory between sessions, specifically to prevent chatbots from building on harmful mental states over time.
Del Rosario eventually came up with his own approach. After the divorce kept leaking into unrelated conversations, he started dedicating separate chatbots to separate parts of his life and using anonymous mode for anything sensitive.
He still values it when it works, like when it knows his children need car seats on a road trip, or when it reminds him he has a lot on his plate. His mom died two years ago, and between that, the divorce, the children and work, the chatbot is sometimes the only thing that gets the full picture.
“It feels good to be seen, even if it is by an AI chatbot,” he says.
Major AI assistants let users turn off memory entirely, and each offers some way to view, edit or delete what’s been stored. But many users don’t know about these newer capabilities, or never check these settings. Here are some tips for making sure the memory feature works to your advantage:
Learn what your chatbot knows about you. Go to your settings—in ChatGPT it is under Personalization, in Claude and Gemini it is in the memory section—and review what’s there and delete if necessary.
Use temporary chats for anything sensitive. Or just don’t tell it anything sensitive. The controls—and names—for temporary chat can differ from app to app, but there is often a button at the top of the page.
Try turning memory off entirely to see how it changes your results. You might find you prefer the trade-off.
Treat it like a social-media profile. It is worth checking in and updating on occasionally, because it is shaping what you see whether you look at it or not.
Write to reports@wsj.com
