Siri AI vs a self-hosted assistant — who owns the memory?

Apple introduced the rebuilt Siri AI at WWDC on June 8, 2026 — a dedicated app, on-screen awareness, personal context drawn from a semantic index of your device, and a developer surface that lets any app expose its data and actions to the assistant. Developer access opened the same day; the public release ships with iOS 27 this autumn.

It is genuinely good work, and it is worth understanding precisely — because the details reveal something more interesting than the marketing does.

How the new Siri actually processes your requests

Three tiers, and which one you hit depends on what you ask.

Tier one — on device. Timers, music, smart-home switches, opening apps. Apple's own models run on the device's neural engine, and the request never leaves the phone. This tier is genuinely private in the strongest sense.

Tier two — Private Cloud Compute. Summarise a document, resolve calendar conflicts, answer a question about your email. These go to Apple Silicon servers running stateless, with no data retention and no privileged runtime access. It is the most rigorous cloud-privacy architecture any major vendor has shipped.

Tier three — world knowledge. Broader questions that require knowledge beyond your device and Apple's own models.

Here is the part that got less stage time: in 2026 Apple partnered with Google and Nvidia, and Apple Intelligence's cloud inference now runs on Nvidia GPUs inside Google's cloud infrastructure, with Private Cloud Compute working alongside Nvidia's confidential computing. The privacy guarantees are still asserted, and the engineering behind them is real. But the physical machine executing your most personal queries belongs to neither you nor, in that tier, Apple.

Security researchers have made the sharper version of this point: private inference isn't the same as private. Even flawless confidential computing means trusting an attestation chain you cannot personally inspect, on hardware you cannot personally reach.

Three constraints nobody mentions in the keynote

Hardware floor. Apple Intelligence needs at least 8 GB of RAM, and the new Siri tiers require iPhone 15 Pro or newer. A large share of the installed base is simply excluded.

Geography. Siri AI is not launching in the EU or China. Mac, Apple Watch and Vision Pro users in the EU can reach it in a supported language, but iPhone users in the world's second-largest premium market are outside the rollout. Meanwhile the European Commission is pushing for rival assistants to get the same prominence and access as Apple's and Google's — and Apple has objected on privacy grounds, arguing it would be forced to give any virtual assistant direct access to users' private data. Whatever your read on that standoff, the practical consequence for a European today is: the assistant Apple advertises isn't the one they get.

Ecosystem gravity. Siri's power comes from App Intents — apps deliberately exposing their content and actions to Apple's assistant. That is a real capability, and it stops precisely at the boundary of the Apple ecosystem. Your assistant can act inside participating apps on your iPhone. It cannot log into a website for you, run a job at three in the morning, or maintain a project across a week.

What a self-hosted assistant does differently

The comparison isn't "better or worse". They are different categories of thing, and confusing them is how people end up disappointed.

Siri is an interface to your device. A self-hosted assistant is a worker with its own infrastructure. Siri answers when addressed. An assistant on your own server runs scheduled work with nobody watching — a morning digest, a weekly report, a task carried across days.

Memory of state vs memory of history. Siri's personal context comes from a semantic index of what is on your device now. A self-hosted assistant accumulates typed long-term memory: facts that get overwritten when they change, timestamped events that never do, corrections that override behaviour, and its own inferences that decay unless reconfirmed. An index tells the assistant what exists. A memory tells it what happened between you.

Where the accumulation lives. This is the whole argument. Siri's heavier processing happens on infrastructure you don't control. A self-hosted assistant keeps the accumulated profile of you — conversations, memory databases, files — in a folder on a VPS that costs a few dollars a month and answers only to you. Model calls still go out per request, but the record of how you think stays home. We've written about what that distinction means in practice.

Reach. No hardware floor, no regional rollout, no app ecosystem boundary. It runs wherever a small Linux box runs, for everyone, including the EU.

So is Apple's move bad news for self-hosting?

Honestly: partly, and it's worth saying plainly. Hundreds of millions of people are about to get a competent assistant that knows their context, for free, already installed. If all someone wanted was a better voice command on their phone, that need is now met, and no self-hosted alternative should pretend otherwise.

But the same launch validates the harder argument. Apple spent years and enormous engineering capital trying to make a cloud assistant you can trust — custom silicon, stateless servers, confidential computing, published attestation. And the honest end state is still: trust our attestation, on our partner's hardware, if you live in the right country and bought the right phone.

That is not a failure on Apple's part. It is the ceiling of the model. Nobody can hand you verifiable control over infrastructure they own, because verifiable control is exactly the thing that stops being theirs the moment they hand it over.

The realistic setup, if you want both

They are not mutually exclusive, and the pragmatic answer is to use each for what it's good at. Let the built-in assistant handle device-level things it does well: timers, playback, home controls, quick on-screen questions. Keep the work that involves your accumulated life — long-running projects, business processes, the memory of what you decided and why — on an assistant you own. You can even bridge the two: a Shortcut on the phone can send a voice request to your own assistant and read the answer back, which makes the built-in layer a microphone rather than a destination.

The dividing line isn't capability. It's whether a given piece of your life should live somewhere you can point at.

FAQ

Is the new Siri private?
Substantially more than most cloud assistants: simple requests never leave the device, and heavier ones run on stateless Private Cloud Compute. But in 2026 Apple's cloud inference runs on Nvidia GPUs inside Google's cloud infrastructure, so the guarantee is an engineering promise about hardware you cannot inspect.

Why is Siri AI not available in the EU?
Apple is not launching Siri AI on iPhone in the EU or China at rollout. EU users on Mac, Apple Watch and Vision Pro can access it in a supported language. Apple has separately objected to EU requirements that would compel giving any virtual assistant direct access to users' private data.

Do I need a new iPhone?
Apple Intelligence requires at least 8 GB of RAM, and the new Siri tiers target iPhone 15 Pro and newer.

When is Siri AI coming out?
Developer access to the rebuilt Siri AI opened on June 8, 2026, alongside iOS 27, iPadOS 27, macOS 27 and visionOS 27. A public beta follows later in 2026, with the full release tied to the autumn launch of iOS 27.

What are the alternatives to Siri AI?
If you are outside the supported regions, on older hardware, or simply want a different trade-off, the realistic Siri alternatives split in two. Cloud assistants — ChatGPT, Gemini, Claude — give you frontier models and keep your history on their servers. Self-hosted assistants run on hardware you own: the assistant layer, its memory and your conversation archive live on your machine, while model calls go out per request. The second category is the only one where the accumulated record of you is not somebody else\'s asset.

Can a self-hosted assistant replace Siri?
Not for device control — it has no OS-level access to your phone. It replaces the layer Siri cannot reach: persistent memory, autonomous scheduled work, integrations beyond the Apple ecosystem, and full ownership of the data it accumulates about you.

What does self-hosting actually cost?
A small VPS in the $5–15/month range is enough, because the server holds state, not inference — the language model still arrives over an API.