Your Second Brain Has a Landlord
Your AI Remembers Everything About You, and Someone Else Owns the Memory
Your assistant now remembers you. Your fears and your finances, your health and your relationships, the questions you only ask at 2 a.m. Quietly, it has become the most intimate dossier ever assembled on a single person, and you do not own it.
The industry calls it a “second brain.” It is worth being precise about the arrangement. You do not own this brain. You rent it. The lease terms are written by whoever runs the model, and they reserve the right to change them.
Persistent memory shipped across every major assistant
2024 → 2025: four of the largest assistants turned persistent
- Feb 13, 2024ChatGPT Memory announcedGA Sep 2024; expanded Apr 2025 to all past chats
- Jul 2025Microsoft 365 Copilot MemoryGeneral availability
- Aug 13, 2025Gemini ‘Personal Context’On by default (opt-out)
- Oct 2025Claude Memory → Pro & MaxRolled out across 2025
Published by Kymata Labs · Independent Research Institution.
If you’ve told an AI something you wouldn’t tell a colleague, yes.
Think back over the last month. You asked it to read a worrying test result and explain it in plain words. You drafted the message you couldn’t bring yourself to write to someone you love. You pasted in a bank statement, a contract, a diagnosis, a confession. You asked the questions you would never type into a search engine that a human might one day read back to you.
And then, here is the new part, it kept all of it. Not as a transcript you might one day delete, but as memory: a standing model of who you are, updated every time you return. Persistent memory shipped across every major assistant between 2024 and 2026. ChatGPT’s memory went generally available in September 2024 and was expanded in April 2025 to reference all of your past chats. Google’s Gemini introduced “Personal Context” on August 13 2025, on by default. Anthropic’s Claude added memory through 2025, reaching Pro and Max in October. Microsoft’s 365 Copilot Memory went generally available in July 2025. The dossier is no longer a feature you opt into. It is the condition you now live in.
Every confession you typed quietly furnished a profile you will never hold the deed to.
“You can delete the app. You can’t evict the landlord from your own memory.”
Kymata LabsWhat it holds, and who can reach in.
A credit bureau knows your debts. A health insurer knows your claims. An advertiser knows what you bought. None of them knows what you ask when you think no one is listening. Your assistant does, and increasingly it keeps a model of it. Two questions matter more than any feature comparison: what is actually in there, and who can reach it.
What’s actually in the dossier
- Your health: symptoms described, results pasted, conditions you haven’t told family
- Your money: salary, debts, the deal you’re weighing, the job you’re quietly seeking
- Your relationships: the names, the conflicts, the messages drafted and the ones deleted
- Your fears: the 2 a.m. questions you’d never put to another human
- Your work: proprietary documents, half-formed ideas, what you actually think of your colleagues
- The inferences drawn on top: not just what you said, but what the system has concluded about you
Who can reach in
- The provider: to train the model, when the setting is on, and to operate the service
- The model itself: surfacing old memories into new answers, in front of whoever’s watching your screen
- A court: a preservation order can override your deletion settings, as one did for roughly six months in 2025
- A future owner: terms are rewritten, companies are acquired, policies change beneath a profile you can’t take with you
- A breach: every retained record is a record that can leak, and the longer it’s kept, the larger the target
Documented, dated, and mostly on the record.
Three claims sit at the centre of this paper: you don’t own it, you can’t fully delete it, and you can’t take it with you. Each rests on a dated, named fact, and several of them were volunteered by the companies themselves.
Deletion is not erasure, and OpenAI says so
The cleanest evidence comes from the vendor. OpenAI’s own memory documentation states it plainly: “Deleting a chat doesn’t erase its memories; you must delete the memory itself.”These are two distinct switches. Clear the conversation and the derived memory can still remain, the thing the system has concluded about you, sitting in a separate store most users never open. The “delete” button you reach for does less than its name implies.
OpenAI, “Memory and new controls for ChatGPT” (2024).A court can override your deletion settings, and one did
This is the anchor. In The New York Times Co. v. OpenAI (case 1:23-cv-11195 in the Southern District of New York), Magistrate Judge Ona T. Wang, on May 13 2025, ordered OpenAI to preserve all output log data, including deleted chats, Temporary Chats, and API contentthat would normally be purged within 30 days. It covered ChatGPT Free, Plus, Pro and Team plus the API. COO Brad Lightcap described the demand directly: “The New York Times is demanding that we retain even deleted ChatGPT chats and API content that would typically be automatically removed… within 30 days.” The obligation was lifted on September 26 2025. For roughly six months, a third party’s lawsuit overrode the deletion rights of millions of users who had nothing to do with it. The settings did what they promised, and it made no difference.
OpenAI, “Response to The New York Times’ data demands” (2025).Retention windows are expanding by 60×
On August 28 2025, Anthropic updated its consumer terms: chats and Claude Code sessions are used for model training when the setting is on, and data retention extended from 30 days to five years for users who allow training. Users had until October 8 2025to choose. Thirty days to five years is a sixtyfold expansion of how long the most sensitive record about you sits on someone else’s servers, and like most of this, it was set by default and by deadline.
Anthropic, “Updates to Consumer Terms and Privacy Policy” (Aug 28 2025).Most people don’t know any of this
A representative survey of 3,270 UK adults asked the basic questions. 53.8% did not know their data could be used to train a model, and 76.5% were not sure whether they could opt out. The consent exists on paper, but the understanding behind it does not. A lease nobody reads is still a lease, and the terms still bind.
Grosse & Ebert, arXiv:2510.27275 (2025).The default is “on”
When Google introduced Gemini’s “Personal Context” on August 13 2025, it arrived on by default, opt-out rather than opt-in. That distinction is the whole game. Under opt-in, the system remembers you only if you ask; under opt-out, it remembers you unless you find the setting and refuse, and most people never go looking. So the dossier fills itself.
Google, Gemini “Temporary chats and privacy controls” (Aug 13 2025).
Most people don’t know the terms they agreed to
UK adults, on how their AI data is used (n = 3,270)
Memory arrived everywhere at once, and quietly.
For most of the chatbot era, the assistant forgot you the moment you closed the tab. That was a privacy feature disguised as a limitation. Then, in barely two years, it disappeared. ChatGPT memory was announced February 13 2024, went generally available that September, and in April 2025was expanded to draw on all of your past conversations. Microsoft 365 Copilot Memory reached general availability in July 2025. Gemini’s Personal Context landed August 13 2025, on by default. Claude’s memory rolled out across 2025, reaching Pro and Max in October. Four of the largest assistants on earth, all turning persistent, inside roughly eighteen months.
Each step was reasonable on its own terms. Memory makes the tool better, because it stops asking you to repeat yourself, picks up where you left off, and feels like it knows you. That is the trade, and it is a real one. But “it knows you” and “someone else owns what it knows about you” are the same sentence read from two ends. You keep the convenience while they keep the dossier.
The lease quietly grew longer, and nobody thought to ask the tenant.
Two kinds of users are forming. Only one read the lease.
On one side are the people who treat memory as a deliberate choice. They open the settings, decide what the assistant gets to keep, know that deleting a chat is not deleting the memory, and understand they cannot port the derived profile to a rival. The tool serves them, and they remain the landlord of their own mind.
On the other side, which is the great majority on the numbers, are the people who never looked. Defaults chose for them, and the dossier filled itself. They believe a deleted chat is gone, assume they could leave whenever they liked, and have never once read what they agreed to. The risk is not evenly shared. It falls almost entirely on the second group, who are also the least likely to feel it coming, right up until a breach, a subpoena, an acquisition, or a quietly rewritten term turns the convenience into an exposure.
The same lease, read by three different readers.
Ownership is not lost, but it is not the default either. The same facts that name the problem also point at the response. What that looks like depends on who you are.
For individuals
Read the memory settings as a lease, and sign deliberately.
- Open memory controls on every assistant and check the defaults, since Gemini’s Personal Context arrived on.
- Delete the memory, not just the chat; they are separate switches, by OpenAI’s own admission.
- Export your conversation history while you can, and treat anything you tell an AI as potentially kept for years, not days.
For employers
Your staff are pasting your secrets into someone else’s memory.
- Assume proprietary material entered into a consumer assistant may be retained for up to five years where training is allowed.
- A preservation order can freeze deletion for months; your data-retention policy is not the only one that applies.
- Provide a governed, enterprise-grade option so people aren’t furnishing a third party’s dossier with your business.
For policymakers
Consent that nobody understands is not consent.
- Three in four adults aren’t sure they can opt out. Mandate plain-language memory terms and opt-in defaults.
- Require genuine deletion of the derived profile, not just the transcript, alongside a real right to be forgotten.
- Create a portability standard so a second brain can move between providers, or accept that switching costs lock people in.
FAQ
You can delete conversations, and you should be able to delete stored memories. But deleting a chat does not delete the memory derived from it. In OpenAI's own words, "Deleting a chat doesn't erase its memories; you must delete the memory itself." Those are two separate switches, and most people only ever touch the first. The derived profile, meaning what the system has concluded about you, persists unless you go and remove it deliberately.
You can export your conversation history, meaning the transcript of what you typed and what the model replied. What you cannot do is export or port the derived memory profile, or move your memory across providers. The transcript is yours to download. The model of you, assembled from those transcripts and the inferences drawn on top of them, stays with whoever runs the model. No portability standard exists for a second brain, so leaving means starting over from zero.
Partly, and pay attention to the defaults. Google's Gemini Personal Context arrived on by default in August 2025 (opt-out rather than opt-in), which means the safe state was the one you had to go and choose for yourself. Read what is on before you assume it is off.
Usually. But "usually" is doing a lot of work. For roughly six months in 2025, a court preservation order in the New York Times litigation required OpenAI to retain output log data that would normally be purged within 30 days, including deleted chats, Temporary Chats, and API content, for ChatGPT Free, Plus, Pro and Team users and the API. The order was entered May 13 2025 and lifted Sep 26 2025. Your deletion settings did exactly what they promised, and a third party's lawsuit overrode them anyway.
It is, and that is precisely the problem, because almost nobody knows what they agreed to. In a representative survey of 3,270 UK adults, 53.8% did not know their data could be used to train a model, and 76.5% were not sure whether they could opt out. Consent that nobody understands is consent in name only.
Open the memory settings on every assistant you use and read them as a lease: what is retained, for how long, whether it trains the model, and how to actually delete the derived profile rather than just the chat. Then decide, deliberately, what this landlord gets to keep. The defaults were not written for you.
Memory is a gift. Just know whose house it’s kept in.
None of this is an argument against the tools. An assistant that remembers you is genuinely better, and pretending otherwise is its own kind of foolishness. It is an argument for knowing the arrangement you’re in. You are renting the most intimate record ever assembled about you, from a landlord who writes the lease, can be compelled to keep what you deleted, and may sell the building. Read the lease. Choose what stays. Don’t mistake a rented mind for your own.
Deleting the app is easy; evicting the landlord from your own memory is the part nobody mentions.
Sources
Every dated claim in this paper is drawn from the primary sources below, most of them published by the companies themselves. Where a fact is a survey finding or a specific court order, we have named it in the text.
- OpenAI (2024). "Memory and new controls for ChatGPT." Memory announced Feb 13 2024, general availability Sep 2024, expanded Apr 2025 to reference all past chats. Includes the company's own statement that deleting a chat does not erase its memories. https://openai.com/index/memory-and-new-controls-for-chatgpt/
- Google (2025). "Temporary chats and privacy controls in the Gemini app." Personal Context introduced Aug 13 2025, on by default (opt-out). https://blog.google/products-and-platforms/products/gemini/temporary-chats-privacy-controls/
- Anthropic (2025). "Using memory with Claude." Memory reached Team and Enterprise around Sep 2025, then Pro and Max in October 2025. https://support.claude.com/en/articles/11817273
- Microsoft (2025). "Introducing Copilot Memory." Microsoft 365 Copilot Memory reached general availability July 2025. https://techcommunity.microsoft.com/blog/microsoft365copilotblog/introducing-copilot-memory-a-more-productive-and-personalized-ai-for-the-way-you/4432059
- OpenAI (2025). "Response to The New York Times' data demands." On the May 13 2025 preservation order in The New York Times Co. v. OpenAI, No. 1:23-cv-11195 (S.D.N.Y.), before Magistrate Judge Ona T. Wang; obligation lifted Sep 26 2025. Includes COO Brad Lightcap's statement on retention of deleted chats and API content. https://openai.com/index/response-to-nyt-data-demands/
- Grosse, M. & Ebert, T. (2025). "What do people know about how their data is used to train AI?" Representative survey of 3,270 UK adults. arXiv:2510.27275. https://arxiv.org/html/2510.27275v1
- Anthropic (2025). "Updates to Consumer Terms and Privacy Policy." Aug 28 2025: consumer chats and Claude Code sessions used for model training when the setting is on; retention extended from 30 days to five years; decision deadline Oct 8 2025. https://www.anthropic.com/news/updates-to-our-consumer-terms
Read it as a PDF
The complete argument, the timeline, and all primary sources, typeset as a white paper. Free, no signup.
↓ Download the PDF