what happens to your data when an ai app shuts down

on december 5, 2025, rewind users found out their app was over. meta had acquired the company, and an update shipped as a kill switch. all screen and audio capture stopped permanently on december 19, 2025.

that is 14 days. two weeks to pull two years of your working life out of a product that had spent those two years telling you the archive was private and yours.

people in the eu, uk, brazil, china, israel, south korea and turkey did not get the 14 days. they lost access immediately.

this post is about what actually happens to an archive when an ai app shuts down, why the pattern repeats, and the four questions worth asking before you hand any tool your history. everything below is current as of september 2026.

why does this keep happening to ai apps?

the category is unusually prone to it, for structural reasons rather than bad intentions.

ai has become genuinely load bearing at work. per the anthropic economic index, 43.36% of classified claude conversations are work, and 51.38% are augmentation rather than automation, which means long sessions where a person and a model work a problem together. that is a lot of thinking to lose in a 14 day window.

it helps to be concrete about what an archive holds. in the same dataset, the things ai conversations produce break down like this.

roughly 30% of that is durable work product, documents, reports and analyses, sitting inside a service somebody else can switch off.

memory products are expensive to run and slow to monetize, which makes them acquisition bait. and the thing that makes them valuable to an acquirer is rarely the product you use. meta bought limitless for the wearables team and technology, not for a mac screen recorder. the founders went to reality labs. the desktop app was overhead, so the desktop app was switched off.

that is the shape of an ai startup acquisition almost every time. the team is the asset, the users are a line item, and the shutdown notice is written by whoever inherited the support inbox.

it is worth being clear that the rewind ai shut down was not unusually badly handled. 14 days of notice and a working export is around the median for this category. that is the problem.

what the privacy policy said, and why it did not help

almost every privacy policy you have agreed to permits this. look for the clause about a merger, acquisition, financing or sale of assets, and it will say your personal information may be among the assets transferred.

that clause is doing enormous work. it means the promises made to you were made by an entity that can stop existing, and the protections you evaluated were the protections of a company that is no longer the one holding your files.

regulatory pressure does not close the gap either. microsoft spent 2 years reworking recall after the privacy backlash of 2024, shipping it off by default with biometric authentication and local storage, and security researchers still rate it a risk. good policy and good defaults are worth having. neither of them is ownership.

so reading the privacy policy tells you how a company behaves today. it tells you very little about who will hold the data in three years. that is a different question, and it is answered by architecture rather than by policy.

the only question that survives an acquisition

if the company vanished tonight, could you still open your archive tomorrow?

that single question cuts through everything else, because it is about where the bytes live and what format they are in. no clause can change the answer.

there are three broad answers, and they behave very differently in a shutdown.

1. cloud only. your data lives on their servers, in their schema. a shutdown means an export window, if you are lucky, in whatever format they choose. rewind's kill switch is this case.
2. cloud with local cache. the working copy is on your machine but the format is proprietary and depends on their service to be readable. slightly better. you have the bytes and no way to read them.
3. local first. the archive lives on your disk in a format that opens without the vendor. a shutdown is an inconvenience rather than an extinction event.

only the third one survives. this is the whole argument for local first software, and it is not really a privacy argument. it is a continuity argument that happens to have privacy as a side effect.

how to export your data before you need to

do this while nothing is wrong. export windows are short, badly timed, and announced by email you will not read.

• find the export today. if you cannot locate it in settings in under two minutes, that is the finding.
• take the most open format offered. sqlite, json, markdown and plain media files beat a proprietary bundle.
• open the export in something else. an export you have never opened is not a backup, it is a hope.
• store it somewhere you control. your own disk or your own bucket, not the vendor's.
• repeat quarterly. data portability is a habit, not a feature. 4 exports a year costs you about 20 minutes.

for screen memory tools specifically, check whether the recordings and the index export separately. a video archive with no searchable index is a hard drive full of nothing you can find, which we get into in how to find something you saw on your screen.

does local first mean nothing ever leaves your machine?

no, and be suspicious of any tool that says it does.

almost every real product touches a network somewhere: sign in, crash reporting, an update check, a model call. the useful claim is narrower and checkable. where does the archive live, and who can read it.

for remynd the answer is that capture, ocr and storage stay on your mac by default, and you can exclude specific apps or sites from capture entirely. it also signs you in, reports crashes, and calls a cloud model when you ask the agent a question. we say so on the security page rather than claiming a purity nobody in this category actually has. the same distinction runs through private ai on your mac.

that is the honest version of own your data. the archive is on your disk, in your control, and it does not stop working because a term sheet got signed.

four questions before you trust the next one

ask these of any memory, notes or recording tool, including ours.

1. where is the archive physically stored? if the answer is "our cloud," everything below is moot.
2. what format is it in, and does it open without you? sqlite and plain files are a good sign. an opaque bundle is not.
3. what is the export, and have you tested it? not whether one exists. whether you have run it.
4. who is the acquirer likely to be, and what do they sell? an ad company buying a memory company is a specific kind of risk, and it is not a hypothetical one any more.

rewind was a good product built by serious people, and none of that mattered on december 19, 2025. the lesson is not to distrust small companies. it is to keep the copy that does not depend on any of them still being here.

we wrote a fuller comparison of where rewind users have gone since in best rewind ai alternatives in 2026.

download remynd for mac and keep the archive on your side of the line.