Microsoft Recall and Apple’s new Siri AI are trying to solve a similar problem.
At some point we have all had the same thought:
I saw something. I read something. I saved something. Where was it?
That is a perfect use case for AI. Not because AI can generate another paragraph of text, but because it can help search messy personal context: messages, screenshots, emails, photos, notes, webpages, files, apps.
In that sense, Recall and Siri AI rhyme.
Both are about building a semantic memory layer. Both involve OCR. Both involve image understanding. Both are trying to make computers searchable by meaning rather than just filenames, folders, and exact keywords.
But the difference is where that memory comes from.
Microsoft Recall started from the screen.
The controversial version of Recall created a searchable history by taking background snapshots of what was on your display every few seconds. Microsoft’s pitch was that this all happened locally, but the product shape still felt wrong. It was not just indexing your computer. It was watching your computer and creating a new database from whatever happened to pass across the screen.
Microsoft has since reworked Recall around opt-in setup, stronger authentication, local protections, and filtering. But the initial reaction was revealing because the product shape had already triggered the deeper concern.
That distinction matters.
A screen is messy. It contains everything: private chats, bank details, health information, work documents, half-written messages, passwords you briefly revealed, things someone else sent you, things you never intentionally saved. Even if the database is local, encrypted, and protected, the capture model feels too broad.
Local surveillance is still surveillance-shaped.
Apple’s approach looks similar on the surface, but the product shape is different.
Siri AI appears to work from a semantic index of things already present on the device: photos, screenshots you saved, messages, mail, app entities, files, and structured content exposed by apps. Apple is also using OCR and image understanding. That part is not magically different.
The important difference is that Apple is not auto-capturing the screen every few seconds to build the index.
A screenshot in Photos is something I chose to take. A message is something already in Messages. A photo is already in my library. An app entity is something the app has deliberately exposed to the system. Visual Intelligence can understand what is on screen, but it is framed as an explicit action, not a passive background recorder.
That changes the trust model.
There is a more speculative version of this question coming next.
Apple is reportedly working on AirPods with cameras. Not cameras in the normal sense. Not little GoPros for your ears. The rumour is that they would be low-resolution sensors used to give Siri more context about the world around you.
That sounds strange until you put it next to Visual Intelligence.
Right now, Visual Intelligence is mostly something you invoke with a phone, a screen, or a headset. You point the camera at something, take an action, and ask the system to understand it. AirPods with cameras would move that input closer to ambient computing. Siri would not just know what is in your messages, photos, mail, files, and apps. It might also know something about the room you are standing in.
That is where the distinction becomes even more important.
If the cameras are used as an explicit input — ask Siri what this sign says, ask where the entrance is, ask what ingredient I am looking at, ask how to get to the next platform — then this is basically Visual Intelligence without needing to hold up a phone.
If the cameras become a passive recorder, it becomes Recall for the physical world.
I do not think Apple will expose these cameras like normal cameras, if they ship at all. My guess is that they would be treated more like a system sensor: available to Siri and Apple Intelligence for narrow contextual tasks, but not available as a raw video feed for every app. That would fit Apple’s usual privacy model. It would also be the only version that makes social sense.
Because cameras on AirPods have an obvious trust problem.
With smart glasses, at least the camera is visible. People can see the frame. They understand, roughly, that glasses might be looking outward. AirPods are different. They are already everywhere. They are small, casual, and socially invisible. Adding cameras to them changes what the product means, even if the cameras are low resolution and even if the processing is local.
The useful version is not:
We will remember everything you looked at.
It is:
When you ask, we can use the world around you as context.
That could be genuinely helpful. Directions that understand landmarks. Accessibility features that describe signs or objects. Shopping help. Cooking help. Finding the right platform in a station. Understanding a poster, a menu, a receipt, or a product label without taking out your phone.
There is also a longer-term spatial computing story here.
AirPods with outward-facing sensors could help Apple learn how to build contextual AI wearables before the smart glasses arrive. They might help with environmental awareness, rough positioning, spatial audio, or hand and object detection. But I would be careful not to overstate the AR angle. Earbuds are not glasses. Their view of the world is awkward, side-mounted, and easily blocked by hair, hats, or head position. For real augmented reality, glasses still make more sense.
That may be the actual roadmap.
First, make Siri understand your private context.
Then, make Siri understand what is on your screen.
Then, make Siri understand what is in front of you.
Eventually, put that into glasses.
The first Apple smart glasses are rumoured to be less like Vision Pro and more like audio-first AI glasses: cameras, microphones, speakers, Siri, but no display. A later version may add a waveguide display and become more recognisably augmented reality. That split makes sense. Before Apple can put pixels in front of your eyes, it needs the assistant to be useful without the pixels.
And that brings the privacy question back to the original point.
The problem is not whether an AI system has memory. A useful assistant probably needs some memory. The problem is whether the memory comes from things you gave it, things your apps intentionally exposed, things you explicitly asked it to inspect, or things it silently captured just in case they might become useful later.
That line will matter even more when the sensors leave the phone and move onto the body.
The Recall model says:
We will remember what you saw.
The Apple model says:
We will help you find and act on what is already yours.
That is why Apple’s version feels more likely to work.
Not because Apple is incapable of building creepy software. Not because “on-device” automatically solves privacy. And not because OCR inside Photos or Spotlight is new.
It works because the default mental model is less alarming.
The Mac and iPhone already index my files, photos, mail, messages, and app content. Making that index semantic is a natural evolution of Spotlight. Letting Siri search it with a local model feels like an upgrade to the operating system.
Recall felt like a second observer sitting behind the glass.
This is probably the lesson for personal AI products.
The winning version of AI memory will not be the one that captures the most. It will be the one that captures the least while still being useful.
A useful assistant needs context, but context has to come from the right places. Existing data. Explicit actions. App-level semantics. Clear attribution. Local processing where possible. Cloud processing only when necessary.
Apple’s approach is not perfect, and it deserves scrutiny. But it understands something Microsoft seemed to miss with Recall’s launch:
People do not just care where their data is processed.
They care how it got there in the first place.