Meta Wrote the Kindest Death Policy in Technology, for the Wrong Half of Your Account

Meta says what the assistant keeps, and the saying is more specific than the coverage suggests.

On 27 January 2025 Meta announced that Meta AI would "remember certain details that you share with it in 1:1 chats on WhatsApp and Messenger", and would "only remember certain things you tell it in 1:1 conversations (not group chats)". The family group where you called the assistant in to settle an argument about a film sits outside that. The private thread where you typed that your mother has started forgetting names does not.

The second record is not a record in any form you would recognise. That same announcement describes the assistant drawing on account information to make answers more relevant, and gives two examples: the home location listed on your Facebook profile, and reels you recently viewed. When the standalone Meta AI app launched on 29 April 2025, Meta said the assistant "delivers more relevant answers to your questions by drawing on information you've already chosen to share on Meta products, like your profile, and content you like or engage with", and that where Facebook and Instagram sit in the same Accounts Center, "Meta AI can draw from both".

Nobody types out what they are quietly drawn to. They watch, and the watching counts.

The mechanics, plainly, because families deserve them before they need them:

• When you die, your Meta AI memories keep existing and nobody in your family can read, request or correct them. Meta publishes no deceased-user route for assistant memory: memorialisation and account removal act on profiles, and as at 29 July 2026 there is no equivalent form, appeal or nominated person for Meta AI chats.

• Meta AI remembers one to one chats, not group chats. Meta announced on 27 January 2025 that the assistant would "remember certain details that you share with it in 1:1 chats on WhatsApp and Messenger", and would "only remember certain things you tell it in 1:1 conversations (not group chats)".

• The second layer is inference, not a list of sentences. Meta describes the assistant personalising answers from account information, with the home location on your Facebook profile and reels you recently viewed as its own worked examples, and draws from both Facebook and Instagram where they share an Accounts Center.

• You can delete memories while you are alive. Meta's January 2025 announcement states that you can delete its memories at any time. That control belongs to the account holder and to nobody else.

• A Facebook legacy contact does not reach the assistant. A legacy contact can pin a post, change the profile and cover photos, and request removal of the account. A legacy contact cannot log in and cannot read your messages.

• Deletion is a request rather than a guarantee. Meta's AI Terms of Service state that "deleting individual messages, entire threads, or even your account may not delete our copy of your personal information".

• Availability varies by market. Memory and personalised responses were announced for the US and Canada in the January and April 2025 posts, and Meta's 24 July 2026 update described newer capabilities rolling out in select markets first.

Meta's 24 July 2026 update, announcing agentic features running on Muse Spark 1.1, describes an assistant that "knows your context" and can "learn your style". The direction of travel is more of you inside the model, not less.

Give Meta the credit first, because the credit is real. Facebook's answer to death is the best one any platform of that size has shipped, and it predates the assistant by years.

You nominate a legacy contact while you are alive. After your account is memorialised, that person can write a pinned post at the top of your profile, update your profile picture and cover photo, and request that the account be removed. The word "Remembering" appears next to your name. What you shared stays visible to the audience you shared it with. Facebook is equally clear about the limits: a legacy contact cannot log in to your account, cannot read your messages, and cannot remove friends or send friend requests. Nobody can log in to a memorialised account.

Those limits are not failures of nerve. They are a company deciding that dying does not make your private messages public property, which is the right call and a braver one than it looks.

Instagram has half of the same system. Anyone can ask Instagram to memorialise the profile of someone who has died, and immediate family can request removal. Instagram states plainly that it cannot provide login information for a memorialised profile, and that logging in to another person's profile is against its policies. There is no Instagram equivalent of the legacy contact you set up in advance.

Now put the assistant in the same table.

The first two rows are what give the third column its force. A legacy contact is a named human being with a defined and deliberately narrow job. The assistant layer has no named human being and no defined job, because Meta built its death features for profiles years before the same company grew an assistant inside them.

That is a predictable gap rather than a scandal. It is also an obligation. Any company that grows a confessional inside a photo album takes on a duty to the family of the person confessing, whether the product was designed for one or not, and the defence that this was only ever a search box will not survive the first time a widow asks in public.

For the profile side of all this, what happens to your Facebook account when you die walks through the memorialisation and removal forms and the documents each one wants. For the only large-platform feature that hands over a whole account rather than a single profile, read how Google's Inactive Account Manager works before you next open your will.

Three things happen to your Meta AI memories after you die, and one of the three is under anyone's control.

They keep existing. Nothing Meta publishes triggers deletion of assistant memories on death, because Meta has no way of knowing you have died until somebody tells the company, and the telling is a profile process. Memorialising a Facebook profile is an action on that profile.

Deleting the account may not clear them. Meta's own AI Terms of Service say so: "Depending on how you use AIs, deleting individual messages, entire threads, or even your account may not delete our copy of your personal information." The same terms say that when information is shared with AIs, "the AIs will sometimes retain and use that information", and advise users not to share anything "that you don't want the AIs to use and retain, such as information about sensitive topics".

Read that sentence as a widow rather than as a user. The single route your family has, requesting removal of the account, is the route Meta's own terms say may not remove Meta's copy.

Nobody in your family can read them, ask about them, or correct them. No inheritance path. No nominated person. No appeal.

A Facebook legacy contact can change the photograph on a memorialised profile and cannot read one sentence you ever wrote. For the assistant that knows what you are worried about there is no nominated person, no request form and no read access, and Meta's own terms say deleting the account may not delete Meta's copy.

Two numbers belong in this section and cannot go there, and the reason is not modesty. Meta publishes no retention schedule for assistant memories, so nobody outside the company can tell you how long a specific memory survives after an account closes. Meta publishes no deceased-user process for the assistant layer either. Anyone quoting you a confident figure for the first is guessing, and anyone describing a form for the second has not read the help pages. Where the published record stops, this page says so rather than filling the gap.

Rights over a dead person's data also differ by country and by state. None of this is legal advice, only a description of what one company's published terms and help pages said on 29 July 2026.

Meta does offer an export. Under Manage your information on Meta AI you can export your information, and I would rather your family had that file than not.

The file is not the answer, for two reasons.

The first is mechanical. The layer that makes the answers feel specific to you is not a set of saved sentences. Meta's own description is of information already spread across your profile, your engagement and your linked accounts, which is inference rather than a document. Nothing is sitting there waiting to be zipped, so a perfect export of every message you ever sent still leaves out the part that made the assistant feel like it knew you.

The second reason is the one families actually run into. A chat log is a record of a person asking for things. Roast chicken timings. A rewritten email. A parking fine. Almost nothing anybody has ever asked an assistant is what a family would want to hear from them in 2049. Your daughter, at fifty-one, on the morning she has to make a decision you would have had a firm opinion about, is not going to open a transcript of you asking about parking fines.

I run a company in this business, so let me state the limit rather than sell past the limit. A system built from material you chose to give can be asked a question, and being able to ask is a real advance over a folder of files. Being able to ask is not the same as the person being there to answer. Any company that blurs those two things will earn the backlash coming for them.

The useful question is not how to get the transcripts out. The useful question is whether anything you would want passed on was ever typed into that box at all. Usually not, and underneath that sits a worry I hear constantly, which is the feeling that you are not important enough to leave anything behind. Underneath every one of these assistant pages sits the harder question: who actually controls your identity after you die.

Start building your legacy Free build, 50 memories, no card required.

Meta is not the villain of this story. Meta is the clearest case of a pattern that runs through the entire category.

All seven of the big assistants will export your data to you. Not one will hand anything to a person you named. One of the seven, Google, lets you nominate somebody at all, and only through a feature that covers a whole account rather than anything built for the assistant.

Every row was checked against the vendor's own published policy on 29 July 2026. These policies change without notice, so treat the date as the important column.

If you use more than one assistant, what happens to your ChatGPT memories when you die is the hub page for the shared shape of the problem. Come back here for the part that belongs to Meta alone, which is an assistant that grew inside apps you joined for entirely different reasons, sitting one menu away from a death policy that was already working.

None of this requires buying anything.

1. Ask the assistant what it holds on you. Open a one to one chat with Meta AI in Messenger or WhatsApp, ask directly what it remembers about you, then ask it to forget one specific thing. Two minutes there will tell you more than any article, this one included.

1. Delete what you would not want stored. Meta's January 2025 announcement confirms memories can be deleted at any time. Do that for the sensitive ones now, while you are the only person who can.

1. Stop feeding the layer you cannot bequeath. Meta's AI Terms advise against sharing information you do not want retained, including sensitive topics. Take the advice at face value.

1. Set a Facebook legacy contact. About a minute of work, and the one Meta-side control that names an actual human being. Facebook's help centre page on legacy contacts, linked in the sources below, carries the current path through settings.

1. Write one line into your will file or your just-in-case note. Something as blunt as: "My Meta AI memories are not covered by my legacy contact and cannot be recovered by anyone." Your executor will otherwise spend a week finding that out.

1. Then answer the question underneath all of this. If a company you never signed up with knows what you are worried about and your family knows none of it, the fix is saying those things out loud somewhere they can be kept, not filing an export request.

Five of those steps protect you. The sixth leaves something behind, and Meta ships no button for the sixth.

Building your own record with us is free: 50 memories, no card at any stage, and a free build that never expires, so walking away for six months and coming back costs you nothing. If you want to know what sits beyond the free build, the Afterlife.ai plans are there when you are ready.

Afterlife.ai™ exists because the layer that models a person should be the layer that person chose, released to people they named.

A guide called Idy interviews you across sessions, in conversation rather than a form, going after the material nobody thinks to record: how you actually make a decision, what you believe about money when you are being honest, the story about your father you have told out loud a hundred times and never once written down. What we build out of those sessions is what we call Personas.

Yours speaks in your own cloned voice, built from your own recordings, and answers from what you actually said. Where you never covered something, Personas do not invent the missing part.

You nominate the people who get access, by name, while you are alive and able to choose, rather than your daughter filling in a form with a death certificate attached and hoping somebody reads it kindly.

Release runs through Executor Lock™. Somebody you named starts the process, evidence of death is verified, and a seven day window runs before anything opens, so a mistake or a bad actor can be stopped inside that window. At the moment the lock triggers, what you built is a complete snapshot of everything you put in. Nothing is pruned. Nothing is thinned out.

The difference is structural. Meta's assistant learns you as a by-product of you using apps for other reasons, holds what it learns on terms you never negotiated, and offers no way to pass any of that to your children. You build with us on purpose, from material you chose to give, and you name in advance the people who get to open what you built. While you are alive you keep every part of the control: add more, change what is there, delete the lot.

We would rather set out the limits ourselves than have you discover them later. We build this, so we know where the edges are. Personas only know what you actually put in, which means fifty memories give your family something recognisable rather than something complete, and the gaps stay gaps until you fill them. A cloned voice is close and not perfect, and people who knew you well can sometimes hear the seam. We are a young Australian company rather than a hundred year institution, which is why the written material you put in stays exportable rather than locked inside our product. And we can do nothing at all about your Meta AI memories. Nothing on this page changes what Meta holds. What we can offer is a second layer, built on purpose, that your family can actually reach.

If the idea raises questions before any of that appeals, the ethics of AI Personas of the dead is the more careful version of the argument.

Can my family request my Meta AI memories after I die?

There is no published process for this as at 29 July 2026. Meta's memorialisation and account removal routes act on profiles. Meta's AI Terms of Service state that deleting messages, threads or even an account may not delete Meta's copy of personal information, so even the removal route is not a guarantee that assistant data is gone.

Does my Facebook legacy contact cover Meta AI?

No. Facebook's help centre lists what a legacy contact can do: pin a post, change the profile and cover photos, request account removal, and download shared content where enabled. It also lists what they cannot do, including logging in and reading your messages. The assistant's memory of you appears in neither list.

Does Meta AI remember what I said in group chats?

No, according to Meta's January 2025 announcement, which says the assistant will only remember certain things told to it in one to one conversations and not group chats. That is a limit on the memory feature specifically. It is not a statement that group messages are invisible to Meta for every other purpose.

If I delete my Meta account, is the AI memory gone?

Not necessarily. Meta's AI Terms of Service say that depending on how you use AIs, deleting individual messages, entire threads, or even your account may not delete Meta's copy of your personal information. Deleting the account is still worth doing if that is what you want, but treat the deletion as a request rather than a guarantee.

Is Meta AI memory available everywhere?

Meta announced memory and personalised responses for the US and Canada in January and April 2025. Its 24 July 2026 update said newer capabilities were rolling out in select markets with more countries and surfaces to follow. Availability changes without any announcement inside your app, so check your own settings rather than relying on any article, this one included.

What if Afterlife.ai shuts down?

The written material you put in is yours and exportable at any time, so the raw material never sits with one company alone. What each plan includes is set out on the Afterlife.ai plans page.

Is it strange to build a record of yourself on purpose?

Fair question, and a common one. The process is closer to being interviewed properly than to anything technological: somebody asks you good questions and you answer them out loud. The stranger fact is that a company you never signed up with already holds a working model of your preferences while your family holds nothing.

• What happens to your ChatGPT memories when you die, the hub page for this question, useful if you use more than one assistant and want the comparison in one place.

• What happens to your Copilot data when you die, for the same gap inside a Microsoft account that also holds your work files and email.

• What happens to your Grok conversations when you die, the closest parallel case, because Grok also lives inside a social platform you joined for other reasons.

• How Executor Lock releases a legacy to the people you named, for how the verified release and the seven day window actually work.

• How Google's Inactive Account Manager hands over an account, the closest thing any large platform offers to naming somebody in advance.

• What happens to your Facebook account when you die, the profile-side process that sits underneath everything on this page.

• What your child will want to ask you at thirty, for the other half of this, which is not a policy question at all: working out what was actually worth passing on.

Meta holds a working model of your preferences. The people who love you hold nothing, and the second half of that sentence is the half you can change tonight.

That is what the free build is for. Fifty memories is enough to get something that sounds like you and answers in your own words, no card is involved at any point, and the free build never expires, so nothing is taken away if you leave the work alone for a year and come back to it later. Start with one memory, on any evening you have twenty minutes.

Start building your legacy Free build, 50 memories, no card required.

Everything above comes from primary sources: Meta's own newsroom announcements about assistant memory and personalisation, the Meta AI Terms of Service, and the Facebook and Instagram help centre pages on legacy contacts and memorialised accounts. Each was read directly and re-checked on 29 July 2026, and the seven-assistant comparison was built the same way, one vendor policy page at a time.

What is not settled is flagged as such in the text rather than papered over. Meta publishes no retention schedule for assistant memories, no deceased-user process for the assistant layer, and no post-deletion window, so this page describes the absence rather than guessing at what happens inside the company. Market availability of memory and personalisation also changes without announcement, so treat anything about your own country as something to check in your own app.

If you find something here that is wrong or out of date, please tell us through the contact page and we will correct both the page and the checked date.

• Meta Newsroom, "Building Toward a Smarter, More Personalized Assistant", 27 January 2025, updated 9 July 2025. Memory limited to one to one chats on WhatsApp and Messenger, not group chats. Memories deletable at any time. Personalisation using Facebook profile home location and recent reels views. Rollout to the US and Canada. https://about.fb.com/news/2025/01/building-toward-a-smarter-more-personalized-assistant/

• Meta Newsroom, "Introducing the Meta AI App: A New Way to Access Your AI Assistant", 29 April 2025. Assistant draws on profile information and content you engage with, and across Facebook and Instagram where they share an Accounts Center. Personalised responses in the US and Canada. https://about.fb.com/news/2025/04/introducing-meta-ai-app-new-way-access-ai-assistant/

• Meta Newsroom, "Meta AI Doesn't Just Think, It Acts", 24 July 2026. Agentic features on Muse Spark 1.1, assistant that "knows your context" and can "learn your style", rolling out in select markets with more countries and surfaces to follow. https://about.fb.com/news/2026/07/meta-ai-muse-spark-doesnt-just-think-it-acts/

• Meta AI Terms of Service, checked 29 July 2026. Retention and deletion language quoted above. https://www.facebook.com/policies/other-policies/ais-terms

• Facebook Help Centre, "What is a legacy contact and what can they do with my Facebook account?", checked 29 July 2026. https://www.facebook.com/help/1568013990080948

• Facebook Help Centre, memorialised accounts, checked 29 July 2026. https://www.facebook.com/help/103897939701143

• Instagram Help Centre, memorialised accounts, checked 29 July 2026. https://www.facebook.com/help/instagram/264154560391256

• Meta Help Centre, managing your information on Meta AI. Settings, Data and privacy, Manage your information, then export your information. Checked 29 July 2026. https://www.meta.com/help/artificial-intelligence/1771195753735844/

• Microsoft Support, "Conversation history in Microsoft Copilot". Source for the 18 month retention figure in the comparison table. Checked 29 July 2026. https://support.microsoft.com/en-us/microsoft-copilot/conversation-history-in-microsoft-copilot

Platform policy checked: 29 July 2026. Last reviewed: 30 July 2026. This page is on a 30 day recheck cycle because assistant memory policies change without notice.