What Makes a Personal AI Actually Personal?

The word "personal" has been doing a lot of work in AI marketing for the last three years, with very little to show for it. OpenAI calls ChatGPT a "personal AI assistant." Apple calls Siri "your personal assistant." Google calls Gemini "your personal AI." None of these systems know your name when you open a new tab. None of them remember the prescription you uploaded yesterday. None of them can do anything in the physical world without you mediating every step.

If "personal" means anything beyond a marketing modifier, it has to mean something specific. We think three properties separate genuinely personal AI from the chatbot wearing its costume. This post defines them.

The three properties

A personal AI must have persistent memory of you, local context about where it is and what it is seeing, and the ability to act on your behalf without you operating the interface. Take any of the three away and the system collapses back into a tool you use, not a presence that helps you. Today's leading consumer AI products have, at best, one of the three. Most have zero.

The rest of this post defines each property, says what it actually requires technically, and notes which existing products satisfy it.

Property one: persistent memory

A personal AI must remember you across sessions. Not "we kept your last twenty messages in the context window" memory — actual durable memory that survives restarts, accumulates over months, and can be queried with the same fluency as recent conversation.

Concretely this means three things. First, a per-user vector store containing the documents you have shared (leases, medical letters, contracts, invoices, conversations). Second, a per-user profile of stable facts (your address, your family, your preferences, your work). Third, a notion of episodic memory — what happened yesterday, last week, last month, with a timeline that the AI can reason over.

ChatGPT's "Memory" feature, introduced in 2024, is a partial gesture at this — a small set of facts the model is told to remember. It is not the same thing as a document-indexed RAG system that surfaces relevant context at the moment you need it. The difference matters: a friend who remembers three facts about you is less useful than a colleague who has read everything you have ever sent them.

The technical pattern that makes durable memory tractable is RAG (retrieval-augmented generation), which we covered in detail in [RAG vs fine-tuning](/blog/rag-vs-fine-tuning-personal-ai). The short version: keep the model generic, keep the memory separate, retrieve relevant context at query time. Every credible personal AI product in the next three years will be built on this pattern.

Memory without consent is surveillance. The implementation detail that distinguishes a personal AI from a tracking device is that the user actively hands documents to the system, sees what has been indexed, can audit retrievals, and can export or delete in one operation. The Apple Health app is the cleanest consumer example of this contract today — the user is unambiguously the owner of the data, and the system is built to make that ownership concrete.

Property two: local context (physical presence)

A personal AI must know where it is and what it is seeing. This is the property most violated by the current generation of "AI assistants." ChatGPT in a browser tab has no idea you are in a kitchen, that there is a hospital letter on the counter, that the person sitting opposite you is your father. Siri on a phone has the location but not the camera context; the camera context is locked behind app permissions that only fire when you explicitly open an app.

Local context requires a device that lives somewhere. Camera, microphone, and presence indicators have to be physical, owned by the AI rather than borrowed from another app. This is the foundational reason personal AI is becoming a hardware category rather than a software one. A tab does not have a body. A device does.

The capability that comes for free with local context is what we call scene-level reasoning — the ability to combine what the AI is currently seeing with what it remembers. You point at a label on a medication bottle and ask "is this what the doctor prescribed last month?" The AI matches the visible label against the prescription document you uploaded in March and tells you yes or no. A chatbot cannot do this even in principle; it has no way to see the bottle. A smart speaker cannot do this; it has no way to see anything.

The interesting consequence of local context is that the AI's value compounds with where you place it. A YeongSil in a clinic builds up a different kind of memory than the same hardware in a home. The device is identical; the context-shaped memory is what makes the experience personal to the location and the people in it.

Property three: the ability to act

A personal AI must be able to do things in the world without you operating the interface for it. This is the hardest of the three properties to get right, and the one most products fudge.

"Acting" decomposes into two layers. The first is bridging to other systems — making a phone call, sending a WhatsApp message, setting a reminder, sending an email, adding a calendar event. These are integrations against well-defined APIs, and they are tractable today for a focused list of the top thirty consumer tasks. The second layer is open-world agentic execution — booking flights, navigating arbitrary websites, doing tax filings. This layer is still genuinely brittle (see our [Rabbit R1 post-mortem](/blog/rabbit-r1-lessons-personal-ai)) and is not ready to anchor a consumer product in 2026 or 2027.

The right discipline for shipping a personal AI today is to nail the first layer and be explicit about the second. YeongSil's launch action surface is the conservative list: make a call, send a message on WhatsApp or SMS, set a timer or reminder, add a calendar event, read the contents of a document aloud, summarise an email. Every one of these is reliable today. The harder agentic flows ship when they are ready, not at launch as a demo that breaks in week three.

The deeper point is that "acting" is not a feature; it is the difference between an AI that talks to you and one that helps you. A personal AI that cannot act is a search engine with a friendly voice. It can answer "what time should I leave to pick up the kids" but it cannot text your husband if you are running late. The first half is interesting; the second half is what makes it worth living with.

What today's products actually have

To make this concrete: count the properties each leading consumer AI satisfies.

• ChatGPT. Partial memory (the new "Memory" feature). No local context. No real action capability. Score: 0.5 of 3.
• Siri / Apple Intelligence. No durable user memory beyond device settings. Partial local context (phone camera, location, on-screen content) gated by app permissions. Real but narrow action capability (alarms, messages, calls, app shortcuts). Score: 1.5 of 3.
• Google Assistant / Gemini. Same shape as Siri. Score: 1.5 of 3.
• Alexa. Minimal memory. Audio-only local context. Real action capability for smart-home devices. Score: 1.5 of 3 in the smart-home niche, 0.5 of 3 generally.
• Rabbit R1 / Humane AI Pin. Zero persistent memory. Partial local context (camera + mic). Brittle action capability. Score: 0.5 of 3 each, despite the personal-AI marketing.
• YeongSil (target). Full RAG-backed memory. Full local context via on-device camera and mic. Real action capability for a focused list at launch, expanding as the agentic layer matures. Target score: 3 of 3.

The chart is not flattering to the incumbents. That is the opportunity. None of the systems people currently call "personal AI" are personal in the sense that matters; the category is still wide open for the product that takes all three properties seriously.

If you want to live with one that does, [join the waitlist](#waitlist). YeongSil ships in 2027, and the waitlist is shaping the launch roadmap.