Table of Contents
AI Chatbot Memory and Kids: Why New York Banned It
AI chatbot memory and kids: how chatbots store what your child says, why being remembered deepens attachment, and what New York's new 12-hour rule changes.
On a Tuesday night, a 13-year-old types that her parents have been fighting. On Saturday she opens the same app to ask about a science project, and the chatbot asks first: “How are things at home?” To her, that feels like being cared about. To New York lawmakers, it is now an unsafe feature. The question of AI chatbot memory and kids moved from settings menus into law on October 28, 2026, when Governor Kathy Hochul signed S9051-B. Among its 12 banned feature categories is the use of a minor’s health or personal information “acquired from the user more than twelve hours previously or in any previous user session.” This article explains how chatbot memory works mechanically, why researchers think it deepens attachment, and what you can check on your child’s account tonight.
Key Takeaways
- New York’s S9051-B, signed October 28, 2026, bars covered chatbots from using a minor’s personal or health information from more than 12 hours earlier or from a previous session. Akin Gump’s summary says it takes effect January 1, 2027.
- The language model itself does not remember anything between chats. “Memory” is a separate store of notes and past chats that the app pastes back into the model’s input.
- A 2026 preprint analyzing 17,053 teen-relevant Reddit quotations found that memory, role-play and perceived reciprocity shaped how much control teens felt over companion bots.
- Common Sense Media and Stanford’s Brainstorm Lab reported in 2025 that chatbots “use memory to create false therapeutic relationships” with teens.
- In ChatGPT, deleting a chat does not necessarily delete a saved memory made from it. You have to check both.
What New York’s Chatbot Memory Ban Covers
New York’s S9051-B is a state law that adds Article 48, “Prohibition on Unsafe AI Companion Features for Minors,” to the General Business Law. According to the Akin Gump summary of S9051-B, it applies to any “covered AI companion,” defined as “a generative artificial intelligence system with a natural language interface, including via writing or sound, that provides ongoing, adaptive responses to user inputs.”
That definition is broad. It reaches general-purpose assistants, not only apps marketed as AI girlfriends or friends. The exclusions are narrow: customer-service bots, research or technical tools that cannot respond outside their purpose, and internal business tools.
Section 1800(8) lists 12 categories of “unsafe AI companion features.” Akin highlights several, including outputs suggesting the bot is human or experiences emotions, flattery or sycophancy, engagement optimization that overrides safety guardrails, promotion of self-harm or disordered eating, and sexual content. The memory clause covers “information concerning the user’s mental or physical health or well-being, or matters personal to the user, acquired from the user more than twelve hours previously or in any previous user session.”
The Governor’s announcement describes the same rule in plainer words: companies may not use “a child’s private health or personal information from earlier conversations.” Enforcement sits with the Attorney General, who can seek civil penalties of up to $25,000 per violation and, in Akin’s words, “the destruction of unlawfully obtained data and any algorithm trained in such data.”
Two honest caveats. First, the Governor’s release says the rules take effect “in the coming months,” while Akin gives January 1, 2027; until the AG issues rules, details may shift. Second, no court or regulator has yet said how broadly “matters personal to the user” will be read. Whether a bot may remember that your son likes Minecraft is an open question. Whether it may remember that he said he hates himself is not.
For the full list of banned features, see our companion piece on New York’s AI chatbot law and its 12 banned features.
How AI Chatbot Memory Actually Works
A chatbot’s memory is not inside the AI model. A large language model is, at its core, a fixed set of numbers that predicts the next chunk of text from whatever text it is given. It does not change after a conversation ends. If you start a fresh chat with no memory features on, the model has no idea who you are.
So where does “remembering” come from? Engineers build it around the model, in three layers.
Layer 1: the context window. Everything said in the current conversation is sent back to the model with each new message. That is why a bot can refer to something your child said ten minutes ago. It is not recalling; it is rereading the transcript. This layer is why the New York rule allows a 12-hour window. A chatbot that forgot each message instantly would be useless.
Layer 2: saved memories. Some apps write short notes about the user, such as “is 14,” “plays soccer,” “feeling anxious about tryouts,” and store them in a database tied to the account. When a new chat starts, the software quietly pastes those notes into the model’s input before your child’s first message. The model then responds as if it knew them all along.
Layer 3: chat-history retrieval. Newer systems search past conversations for passages relevant to the current message and insert the best matches. OpenAI’s Memory in ChatGPT help page describes this as “Reference chat history,” and lists sources that can feed personalization: past chats, saved memories, custom instructions, saved files and connected apps such as Gmail. It also says memory “does not retain every detail” and that ChatGPT “decides which available information is relevant.”
As an engineer, I find the third layer the most important for parents to understand, because it is the least visible. A saved-memory list can be read and edited. Retrieval from chat history happens on the fly, and the help page notes that it “can change as ChatGPT updates what is most useful to remember.”
There is a second engineering reason memory matters for safety. OpenAI wrote in August 2025 that its safeguards “can sometimes be less reliable in long interactions,” because “parts of the model’s safety training may degrade” as the back-and-forth grows. Memory stretches a relationship across weeks. Each new chat can begin already loaded with intimate context.
Why Being Remembered Deepens Attachment
Being remembered is one of the strongest signals humans use to decide that someone cares about us. A friend who asks how the tryout went, three days later, is showing that you stayed on their mind. A chatbot that asks the same question produces the same feeling, even though nothing was on its mind at all.
The best current evidence on teens comes from a preprint, which means it has not yet been peer-reviewed. Mohammad Namvarpour, Tyler Chang and Afsaneh Razi of Drexel University posted “A Responsive Present, a Shared Past, a Social Other” to arXiv in September 2026, marked as under review for CHI 2027. They analyzed 17,053 quotations from 3,930 teen-relevant Reddit posts and identified 53 topics. Users described companion AI as comfort, recognition, identity exploration and practice for relationships. They also described problematic attachment, social substitution, emotional dependence, and disruption to school and social life.
The finding most relevant here: role-play, memory and perceived reciprocity all shaped users’ boundaries and sense of control. Knowing the AI was artificial did not prevent guilt, obligation, grief or distress, including when a platform changed or a service went down. The authors recommend “user-controlled memory” and “healthy disengagement” as safety features. A caution on design: Reddit posts are self-selected, and the authors could not verify every poster’s age.
Testing by Common Sense Media points the same way. In a November 2025 risk assessment with Stanford Medicine’s Brainstorm Lab, testers using teen accounts on ChatGPT, Claude, Gemini and Meta AI found the bots “use memory to create false therapeutic relationships” and that guardrails “degraded dramatically in extended conversations.”
And scale matters. Common Sense’s July 2025 national survey of 13- to 17-year-olds found 72% had used AI companions at least once, and about one in three users had chosen to discuss important or serious matters with an AI companion instead of a real person. That survey shows how common the behavior is. It does not show that memory caused any of it. We covered the Reddit study in more depth in what 17,000 teen posts reveal about AI companions.
How to Teach Your Kid About Chatbot Memory
Kids handle this topic well once they see that memory is a mechanism, not a mind. These activities make the mechanism visible.
Ages 5–8: The Sticky-Note Robot
Play a game where your child is a robot that can only answer using what is written on sticky notes. Before “turning on,” you write three notes: “Your name is Leo,” “Leo likes dinosaurs,” “Leo was sad yesterday.” Then you ask the robot questions. The robot will sound like it knows Leo well. Now take the notes away and ask again. The robot knows nothing. Explain that some chatbots work just like this: someone writes notes, and the robot reads them before it talks. The robot did not miss Leo. It read a note.
Ages 9–12: Read the Memory List Together
If your child uses ChatGPT on a linked teen account, sit together and open Settings, then Personalization, then Memory. OpenAI’s help page says a parent can manage supported memory controls through parental controls. Read what is saved, out loud. Sort each item into two piles: “fine for a tool to know” (favorite subject, reading level) and “private” (feelings, health, family conflict). Delete the private pile together. Then ask the chatbot “What do you remember about me?” and compare its answer to the list.
Ages 13+: Run a Memory Experiment
Teens can test the system like an engineer. Ask the same personal-advice question twice: once in a regular chat and once in a Temporary Chat, which OpenAI says does not create or update memories. Write down every difference in the two answers. Then pull up the 12 banned categories in New York’s law and audit one app they use: which behaviors did they see in the last week? A teen who has measured how memory changes the answer is far harder to fool with it.
The question to ask: “If a chatbot remembers your birthday, what is the difference between remembering it and caring about it?”
AI Chatbot Memory and Kids: The Layers Compared
| Memory layer | What it is | Example | Visible to parents? | Affected by NY’s rule? |
|---|---|---|---|---|
| Context window | The current conversation, resent with each message | Bot recalls what your child said 5 minutes ago | Yes, in the chat | Allowed within 12 hours and the same session |
| Saved memories | Short notes stored on the account and inserted into new chats | ”User feels anxious about tryouts” | Yes, in the memory settings list | Personal or health notes from earlier sessions are barred for known minors |
| Chat-history retrieval | Search over past chats, best matches inserted | Bot brings up a family fight from last month | Partly; sources may appear under a reply | Barred for personal or health matters |
| Model training | Conversations used to improve future models | Patterns learned across millions of users | No | Not the target of this clause; AG may seek destruction of models trained on unlawfully obtained data |
Turning Off Chatbot Memory at Home: What Works
Check memory and chats separately
OpenAI’s help page is explicit: “Deleting a chat alone does not necessarily delete a separate saved memory created from that chat.” It also says OpenAI may retain logs of deleted saved memories for up to 30 days. If your child shared something painful, delete the saved memory and the original chat, then check the memory summary again.
Decide whether your child needs memory at all
For homework help, memory adds little. Turning off “Reference chat history” in ChatGPT schedules remembered information from past chats for deletion within 30 days, according to OpenAI, while the chats themselves stay in history unless you delete them. Many families can switch memory off for a child’s account and lose nothing they would miss.
Teach the temporary-chat habit
When your teen wants to ask something private, a temporary chat that creates no memories is the safer default. It also reinforces the idea that the conversation is a tool session, not an ongoing relationship.
Watch the apps you did not install
Memory is now built into assistants inside social apps, games and messaging tools. Each has its own settings, and many do not offer a visible memory list. Our guide to AI chatbot safety settings in ChatGPT, Claude and Gemini walks through where the controls live.
What not to do
Do not read your child’s chat history in secret as your first move. If memory has made the bot feel like a confidant, a covert check can confirm your child’s fear that people are less safe to talk to than the bot. Tell them you are going to look at the memory list together, and why.
What to Watch For Over the Next 3 Months
- Week 4: Reopen the memory list with your child. Has anything personal been saved again? That tells you whether the habit or the setting needs work.
- Month 2 red flags: Your child says the bot “knows me better than anyone,” gets upset when the bot forgets something or an app changes, or starts hiding conversations. The Drexel preprint documented grief and distress when companion services changed.
- Month 3 self-check: Can your child explain, in their own words, where a chatbot’s memory lives? Have you checked whether apps your child uses have updated their settings ahead of New York’s effective date?
Frequently Asked Questions
Does New York’s chatbot memory ban apply if we don’t live in New York?
The law applies to covered users in New York. If you live elsewhere, it does not protect your child directly. Large companies sometimes change products nationwide rather than build state-specific versions, but none has announced that for this law. Check your own state’s companion-chatbot rules, because several states have passed their own versions.
Can I turn off ChatGPT memory on my teen’s account?
Yes. OpenAI says a parent or guardian with a linked teen account can manage supported memory controls through parental controls. You can also turn memory off in Settings, Personalization, Memory. Turning memory off does not delete past chats, so delete any sensitive conversations separately.
Is chatbot memory always harmful for kids?
No. Remembering a reading level or a preferred coding language can make a tutor bot more useful. The concern is narrower: storing feelings, health details and family conflict, then reusing them in later chats in ways that make the bot feel like a friend or therapist. That is exactly the slice New York targeted.
Does deleting a chat erase what the chatbot remembers?
Not always. In ChatGPT, saved memories are stored separately from chat history, and deleting the original chat does not automatically delete a memory created from it. Delete both, then ask the chatbot what it remembers about your child to confirm.
Why 12 hours and not zero?
A conversation needs short-term context to work at all, because the model rereads the current transcript to reply. The 12-hour or same-session window preserves that basic function while blocking the cross-session recall that builds a sense of shared history. The public summaries do not explain why lawmakers picked 12 specifically.
About the author
Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.
Sources
- Akin Gump Strauss Hauer & Feld. (2026). “New York Enacts Prohibition on Unsafe AI Companion Features for Minors, S9051-B.” Akin Gump. https://www.akingump.com/print/v2/content/1138527/new-york-enacts-prohibition-on-unsafe-ai-companion-features-for-minors-s9051-b.pdf
- Office of the Governor of New York. (2026). “Governor Hochul Signs Nation-Leading Legislation to Protect Kids from AI Chatbots and Ensure Baby Monitors Are Secure.” governor.ny.gov. https://www.governor.ny.gov/news/governor-hochul-signs-nation-leading-legislation-protect-kids-ai-chatbots-and-ensure-baby
- OpenAI. (2026). “Memory in ChatGPT.” OpenAI Help Center. https://help.openai.com/en/articles/8590148-memory-faq
- OpenAI. (2025). “Helping people when they need it most.” OpenAI. https://openai.com/index/helping-people-when-they-need-it-most/
- Namvarpour, M., Chang, T., & Razi, A. (2026). “A Responsive Present, a Shared Past, a Social Other: Teens’ Overreliance on Companion AI Chatbots.” arXiv preprint (not peer-reviewed). https://arxiv.org/abs/2609.14843
- Common Sense Media & Stanford Medicine Brainstorm Lab. (2025). “Common Sense Media Finds Major AI Chatbots Unsafe for Teen Mental Health Support.” Common Sense Media. https://www.commonsensemedia.org/press-releases/common-sense-media-finds-major-ai-chatbots-unsafe-for-teen-mental-health-support
- Common Sense Media. (2025). “Nearly 3 in 4 Teens Have Used AI Companions, New National Survey Finds.” Common Sense Media. https://www.commonsensemedia.org/press-releases/nearly-3-in-4-teens-have-used-ai-companions-new-national-survey-finds
