Does Meta Keep the Recordings From Your Ray-Ban Glasses?

Its own Notice puts a year on your Voice and no number on the rest
Take a photo with a pair of Ray-Ban Meta glasses and it sits on the glasses until you decide to move it. Say "Hey Meta, what am I looking at?" and, in that instant, a recording of your voice leaves the frames, travels to a Meta data centre, and starts a clock that Meta's own notice sets at one year. The two actions look identical from the outside: the same white capture light blinks for both. They are not the same thing at all, and the difference is the whole answer to a question a lot of new owners only think to ask after the box is open: where does this go, and how long does Meta keep it.
Most of what has been written about these glasses answers that question secondhand, through a privacy group's summary or a reporter's paraphrase. Meta publishes the answer itself, in two documents anyone can open: a Voice Controls Privacy Notice for the glasses, effective 22 July 2025, and the general Privacy Policy that sits under everything the company runs. Read straight through, they say something sharper than the coverage on one point and something vaguer than either side lets on everywhere else.
A year for your voice, ninety days if the glasses woke by accident
The sharp part is voice. When you speak to the glasses after the "Hey Meta" wake word, the notice says the audio recording, the text transcript of that exchange, and related data are stored, and by default they are saved to improve Meta's products. The retention is not open-ended. Recordings are kept up to one year. If the glasses triggered by mistake, heard the wake word in a passing conversation and started listening when you never meant them to, that stray recording is deleted within 90 days of Meta detecting it.
That 90-day figure is worth pausing on, because it is the kind of specific a reader can actually use, and it appears in Meta's own notice rather than in any of the coverage that summarised the policy last spring. It is also a quiet admission: the glasses do wake up on their own often enough that Meta wrote a separate, shorter clock for it.
None of this is the camera roll. The photos and videos you shoot and keep are a different pile, and Meta treats them differently: more on that below. The one-year clock is for the things you say.
The opt-out that used to exist
There was a version of these glasses you could tell to stop. Until the spring of 2025 an owner in the United States could switch off the storage of their voice recordings in Meta's cloud. On 29 April 2025 Meta changed that. TechCrunch, which flagged the update at the time, found that AI features were now on by default and that the switch to keep recordings out of the cloud was gone. As one account of the change put it plainly, users "no longer have the option to disable voice recording entirely."
It is easy to read that as no control at all, and it is not. Two levers remain, and they are not nothing. You can delete recordings by hand, one at a time, from the voice activity log in the companion app; the notice says that once you do, the data is no longer available to be replayed or reviewed and stops being used to improve the product. And you can turn the voice-command feature off altogether: the glasses still take photos, they just stop being something you talk to. What you cannot do any more is keep talking to them and keep the recordings out of Meta's cloud. The default is store; the undo is a chore you perform after the fact or a feature you give up.
An opt-out you have to operate every day, per recording, is a different creature from a setting you flip once. Meta did not remove the ability to delete. It removed the ability to not have anything to delete.
The reviewer at the far end is sometimes a person
The other thing Meta's own words settle is who, or what, looks at the stored audio. The reassuring assumption is a machine: that "trained reviewers" is a euphemism for a model. It is not only a model. Meta's statement, quoted in the coverage of the change, is direct: "Stored audio recordings and related data are used to improve Meta's speech technologies (for example, our speech-to-text technologies) with the help of trained reviewers." The notice itself pairs machine learning with expert reviewers who process the information for product improvement, troubleshooting and training.
This is the same pipeline that turned into a lawsuit and a headline earlier this year, when two owners sued Meta over human review of what the glasses captured. The graphic specifics of what those reviewers reportedly saw are contested and secondhand, and they are not the point here. The point is narrower and comes straight from Meta: agree to talk to these glasses and you agree that a person you will never meet may, at some point, process what you said.
Ask how long Meta keeps everything else, and there is no number
Voice is the part Meta pins down. For the rest, its Privacy Policy is built to avoid a number, and reading it is a useful corrective to the idea that a retention question has a retention answer.
Under the heading "How long do we keep your information?", the policy says Meta keeps information as long as it needs it, and decides how long that is "on a case-by-case basis." It lists what tips the scales: whether it needs the data to operate the product, the feature the information is used for and how that feature works, legal obligations, and other purposes such as safety and security. The one concrete illustration it offers is not about glasses at all: messages sent in Messenger's vanish mode, it notes, are kept for less time than ordinary messages. And it reserves the right, "in some instances," to hold information "for an extended period of time." A searcher who came for a figure leaves with a framework.
That framework is the honest answer for the camera roll. The photos and videos you shoot and never hand to the assistant are, on Meta's own account, stored privately on the glasses and imported to your phone only when you choose — "you, and only you," the company says, "unless you choose to share them." They are not uploaded to Meta's cloud by sitting in your gallery, and they are not used to train anything. That protection holds right up until you use an AI feature, because there is no AI running on the glasses themselves. The moment you ask them to read a sign or identify a plant, the footage is sent to Meta to do it — a boundary the sibling piece to this one, on what Meta may train on, walks in detail. Before that moment, Meta publishes no retention clock for your media because, by its account, it is not holding your media at all.
What the answer actually is
So the real answer to "does Meta keep the recordings from my glasses" is not yes or no. It is: it depends entirely on which recording, and the line between them is not the camera light everyone watches — it is whether you spoke to the glasses or merely looked through them.
Look through them, and the footage stays yours until you move it, on a timetable Meta declines to fix because it says it is not keeping it. Talk to them, and a recording of your voice is in Meta's cloud by default, kept up to a year, processed by machines and sometimes by people, deletable only by hand and only after the fact. The recording light on the front tells the room one true thing — that something is being captured. It says nothing about which of these two fates the capture is headed for, and that is the fate the wearer, not the bystander, has to weigh.
Meta answered the storage question in more detail than it is usually given credit for, and left a hole exactly where the volume is loudest. It will tell you, to the day, how long it keeps the sentence you spoke. It will not tell you how long it keeps anything else, because the answer is a paragraph, not a period of time.
Related coverage
Can the Police Get the Recordings From Your Meta Smart Glasses? — the retention piece asks how long Meta keeps your recordings; this asks who else can compel them from that same cloud
Do Smart Glasses Collect Children's Data? — what the glasses capture and keep, and whether a child's captured data is ever collected by the service
Can Your Smart-Glasses Footage Be Subpoenaed in a Divorce or a Lawsuit? — which copy a civil subpoena can reach turns on the on-device gallery versus the cloud split this piece maps
Sources and verification
- Meta, "Voice Controls Privacy Notice," meta.com/legal/ai-glasses/voice-controls-privacy-notice/, effective 22 July 2025 — voice recordings, transcripts and related data stored and saved by default to improve Meta's products; kept up to one year; accidental activations deleted within 90 days of detection; machine learning and trained reviewers process the data; deletion via the voice activity log ends replay/review and product-improvement use. Read via the page's Japanese localisation (this newsroom's connection is geolocated to Japan); every figure below is corroborated in English by a reachable secondary outlet except the 90-day accidental-activation figure, which is Meta's own. Link
- Meta, Privacy Policy, "How long do we keep your information?", facebook.com/privacy/policy — retention decided "on a case-by-case basis"; factors include operating the product, the feature and how it works (Messenger vanish-mode messages "retained for less time than regular messages"), legal obligations and other legitimate purposes; information kept "for an extended period of time" in some instances. Link
- TechCrunch, "If you own Ray-Ban Meta glasses, you should double-check your privacy settings," 30 April 2025 — the 29 April 2025 change; AI on by default; voice stored after "Hey Meta," kept up to a year; no storage toggle, manual deletion only. Link
- the-decoder, "Meta Ray-Ban smart glasses now record your voice by default to train Meta's AI models," 1 May 2025 — automatic voice recording by default; "Users no longer have the option to disable voice recording entirely"; manual deletion in the companion app; the option to disable voice control; Meta's direct quote on trained reviewers and speech technologies. Link
- eMarketer, "Ray-Ban Smart Glasses users lose opt-out as Meta prioritizes AI training," 1 May 2025 — voice recordings saved in the cloud by default "with no option to opt out"; used for product improvement including AI training. Link
- Electronic Frontier Foundation, "Think Twice Before Buying or Using Meta's Ray-Bans," March 2026 — no local AI, footage fed to Meta when an AI feature is used; "Cloud media" default on; media auto-imported into the required Meta AI app. Link
- Meta, "Meta's AI Glasses: Your Questions Answered," about.fb.com, published 7 July 2026, updated 23 July 2026 — photos/videos "stored privately on your glasses," imported when the wearer chooses; access "You, and only you — unless you choose to share them." Link
- TechCrunch, "Meta confirms it may train its AI on any image you ask Ray-Ban Meta AI to analyze," 2 October 2024 — images/videos shared with Meta AI may be used to improve it, "currently US and Canada"; media never shared is not used. Link
- Meta, "United States Regional Privacy Notice," meta.com/legal/united-states/ — personal information kept "as long as needed" for product delivery, legal compliance and protection of interests; access and deletion via a privacy web form. Read via Japanese localisation, as above. Link
What we could not confirm and left out: a retention period for photos, video or any non-voice media (Meta publishes none, and its policy is case-by-case); that captured media you never share reaches Meta's cloud (Meta and the EFF agree it does not); the graphic specifics of what human reviewers encountered (contested, secondhand); and that the removal of the storage opt-out applies outside the United States, which is where the reporting places it.