What you should demand from any AI that remembers you
AI memory is no longer a feature; it is the default. The frontier assistants now remember you across conversations, in the background, by design, and MIT Technology Review called what AI remembers about you "privacy's next frontier" in January 2026. What has not arrived is a standard: a plain-language list of what you, the person being remembered, are owed. Regulation gives you fragments, GDPR's access, erasure, and portability rights if you are in Europe, and vendor settings pages give you toggles, but nobody hands you the whole demand list. So here is one. Eight things you should demand from any AI that remembers you, whoever makes it, including us. We build Blue, an AI companion whose whole premise is memory, so this page ends by holding Blue to all eight, line by line, including the two places it falls short today. A standard a company will not apply to its own product is marketing, and this page is not that.
A memory bill of rights, in plain language
1. You can see everything it holds about you. Not a sample, not a summary of a summary: the whole picture the system carries of you, laid out where you can read it. The industry's default is the opposite, a partial view over a deeper layer, and the leading assistant's own documentation says so: OpenAI's Memory FAQ states that the memory summary "will not include everything that ChatGPT remembers based on your chats" (their wording, checked August 8, 2026). The discomfort people report about modern AI memory traces exactly to that gap. Demand the whole picture; a partial view fails the standard.
2. Every claim about you cites its source. If a system believes something about you, you should be able to see where the belief came from: which conversation, which words. Beliefs without sources cannot be checked or contested; they can only be trusted or feared.
3. You are the author of the interpretation. This is the deepest one. The doctrine: What people fear about AI memory is usually not the facts; it is the reading, the story the system builds from the facts and the moments it chooses to retell it. The industry's standard answer is deletion-as-hygiene: a console where you weed your own file, entry by entry, because the system has no judgment about which memories still fit you. We think that is the wrong paradigm answering the wrong fear. The right demand is comprehension: a system that holds your history as a story that moves, so that a two-week-old bad night is not engaged as today's news, and a correctable reading, so when the interpretation is wrong you say so and it changes. The forget button is the apology of the paradigm that has nothing between you and the file. Demand authorship of the read, not just scissors for the record.
4. Your memory is portable. You can take it with you, in a form another life can use. Anything less is lock-in wearing a memory costume: the more it knows you, the more it costs to leave, unless leaving with your story is a built-in right.
5. It never trains anyone else's models, and it is never sold. Your memory exists for you. The moment it also serves someone else's model or balance sheet, you are not the customer of your own life record.
6. It survives the company. Products die; October 2025's Dot shutdown proved even the graceful version costs users their continuity. The demand is that the ending, whenever it comes, cannot strand the record: export that works today, not at announcement time. The full checklist for this one is its own page.
7. There are walls between contexts. Memory must not leak across the rooms of your life. The sharpest version is the two-party case: in HeartWeave, our couples product, each partner has a fully private room, and what you say there is never shown to your partner; the shared space is built from what is true from both sides, not from anyone's private words. That wall is the product's founding rule, in the help center. Solo products have the same duty in softer form: what you tell a companion in confidence should never surface in a context that exposes you.
8. Full deletion exists. When you leave, you can take the record (article 4) and end it: account deletion that actually deletes. Not archive, not anonymize-and-keep: deletes.
If a product you use fails several of these, the answer is not necessarily to flee; it is to know exactly what you are trusting and to stop telling it the things you would not put in an unlocked drawer. The audit demand, applied product by product, has a ten-minute checkable form.
Why "just delete memories" is not the answer
Because the console model concedes the fight before it starts. If the only way your memory stays livable is that you personally weed the file, then the system has admitted it cannot tell which parts of your past still describe you, and you have been hired as the unpaid librarian of your own record. Per-entry deletion is a reasonable emergency exit. As a paradigm, it is storage abdicating the actual job, which was understanding. The long version of that argument is in the memory pillar, and the industrial version, storage marketed as memory, is in how AI memory actually works.
Blue, held to the standard, line by line
We grade ourselves below, so check the grades against the five-question test yourself.
1. See everything: yes, by architecture. Blue's understanding of you lives in the Living Profile, synthesized nightly from your conversations, readable in full. Not a sample over a hidden layer; the profile is the layer.
2. Citations: yes. Every facet of the profile cites the conversations it came from. Why Blue believes something about you is a link you can open.
3. Authorship of the interpretation: yes. Weekly, Blue shows you what it learned about you; when the read is wrong, you say "This isn't me" and the profile updates, at the next nightly pass rather than the same instant. And the flip side: there is no button that makes Blue forget an event. You steer the interpretation, not the record; export and full account deletion are the hard exits. If per-entry deletion is your requirement, the console-style systems offer it and Blue deliberately does not. The bet of the whole product is that what people fear is not being remembered; it is being remembered without understanding.
4. Portable: partly today. Export exists and your memory is yours to take. The candid gap: export is currently fulfilled by asking us rather than by a self-serve button, and we would rather say that plainly than hide it behind a settings page that does not exist.
5. Never trained on, never sold: yes. Your conversations are encrypted at rest, never sold, never used to train foundation models. That is a company commitment, in writing, in the help center.
6. Survives the company: what we can say. The export right is how a small company makes this promise credibly, and Blue is a young product from a small company; judge us by the same test as everyone else. What we can say is that the synthesis itself, not just the transcripts, is visible to you all along, so what you would carry out is something you have already seen.
7. Walls: yes, and it is the founding rule of the sibling product. Blue and HeartWeave do not share data, period; walls between products are the company's message, not a settings default. And within HeartWeave, the private room wall is the two-party case above.
8. Full deletion: yes. In-app account deletion, with most data cleared within 30 days.
Six clean, two graded honestly. The gaps are the roadmap.
Who this page is not for
If you want an AI with no memory at all, that is a coherent position and the stateless tools serve it; this standard is for people who have decided memory is worth having and want it on human terms. And none of this is legal advice: if you are exercising GDPR or state-privacy rights against a vendor, the regulator's page beats our essay.
Frequently asked questions
What does AI actually remember about me?
More than the settings page shows, in most products. Modern assistants store explicit facts, plus impressions synthesized from your conversations in the background, and OpenAI's own Memory FAQ says the visible summary "will not include everything that ChatGPT remembers" (checked August 8, 2026). The only general answer is the audit demand above: if you cannot see it all, assume there is more.
Can I see what an AI knows about me?
Partially, in most products: a memories list, a profile summary, or in the EU a GDPR access request for stored data. Full visibility, everything, with sources, is rare by design, and it is the first thing this page says to demand. Products built for auditability do exist; test any claim with the lookup-and-audit questions in the memory test.
Can I make an AI forget something?
In console-style systems, usually yes, entry by entry, through settings. In Blue, no, and we say so rather than blur it: there is no per-memory delete; you correct the interpretation, export everything, or delete the account whole. Which model you prefer is a real choice between paradigms, and this page argues for comprehension over hand-weeding.
What rights do I legally have over AI memory?
In the EU and UK, GDPR gives you access, erasure, and portability rights over personal data, which covers AI memory stores. Several US states have partial equivalents. Everywhere else, and practically even where the law applies, your working rights are the ones the product was built to honor, which is why an architectural standard matters more than a settings page.
Is this the same as the AI Bill of Rights?
No. The White House Blueprint for an AI Bill of Rights (2022) is government policy guidance about automated systems broadly. This page is a user-side standard for one specific, newly urgent thing: products that remember you. Narrower, plainer, and checkable against any app in ten minutes.
Why does Blue not have a forget button?
Because the forget button is the fact-bucket paradigm apologizing for itself. A system that dredges stored fragments needs you to weed them; a system that holds your story as a moving narrative is designed to make the old vent not today's news, by comprehension rather than deletion. Blue's version of control is visibility, citations, correction, export, and full deletion. Where that bet fails on a given week, the misread is visible and correctable rather than silent, which is the whole difference.
If you want to see what memory on these terms feels like, Blue is free to start: activatedhuman.earth/life.