Personal AI

A personal AI that actually knows your work.

The AI assistants most people use are impressive and completely amnesiac about you. They know a great deal about the world and nothing about your clients, your projects, your correspondence or how you decided things last year. A personal assistant — one with access to your own material, running on your own machine — is a different and more useful thing.

What makes it personal

It has read your material. Your documents, notes, emails, past work. Not as training data, but as something it can search and cite when answering.
It remembers. Context that persists between conversations, so you are not re-explaining your situation every time.
It knows your preferences. How you write, what you care about, the decisions you have already made and do not want revisited.
It runs where your data is. Which is what makes all of the above safe to do.

Why the last point is not optional

An assistant is only useful in proportion to what you let it see, and you can only let it see everything if nothing leaves. The two properties are linked: a genuinely useful personal assistant needs access to material you would not paste into a chat window, which means it has to run somewhere you control.

This is the practical case for local AI, distinct from the ideological one. Not privacy for its own sake — privacy as the precondition for usefulness.

How one is actually built

A local model via Ollama on a machine with enough memory — sizing here.

An index over your own files. Documents are split, converted to vectors and stored locally, so relevant passages can be retrieved and given to the model when you ask something. This is what lets it answer from your material rather than inventing.

A memory layer — a simple store of facts and preferences that persist across sessions.

Connections to where you actually work. Your notes, your calendar, your files, your email if you want it. Each connection increases usefulness and the care required.

An interface you will genuinely use. A terminal, a chat window, a keyboard shortcut. Tools that require a detour get abandoned regardless of how good they are.

What it is good at, and what it is not

Works well	Does not
Answering from your own documents	Reasoning at the frontier level
Drafting in your voice	Knowing anything after its training data
Recalling decisions and context	Being reliably correct without checking
Sorting, tagging, summarising	Replacing your judgement

The failure mode to design against is trust without verification. An assistant that knows your files will answer confidently about them, including when it has retrieved the wrong passage. Build it to cite what it used — the source, the file, the passage — so you can check in a second rather than assuming.

Where this is heading

The trajectory is toward assistants that are individual rather than shared: a system that has genuinely accumulated your context, that runs close to you, that does not reset. We think that is where useful AI ends up for most people, and we have written about why at more length in our position on local and personal AI.

Common questions

Can I build a personal AI assistant myself?

The components are available and reasonably approachable: Ollama for the model, an open-source vector store for your documents, and a few hundred lines to connect them. A technically comfortable person can have something useful running in a weekend. Making it reliable, secure and integrated into your daily work is the longer part.

How is this different from just using ChatGPT?

A general assistant knows the world; a personal one knows your work. The difference shows up in what you can ask — "what did we agree with this client in March" is answerable by one and not the other. It also means you can ask about material you would not paste into someone else's system.

Is it safe to give an AI access to all my files?

If it runs locally, the data does not leave your machine, and the remaining risk is who can query the system. That is a real question — an assistant with access to everything is a single point through which everything can be requested. Access control and honest thinking about what to include matter more here than in almost any other setup.

What hardware do I need?

A machine with 16GB of memory runs a small model adequately; 32GB or more is comfortable. For a personal assistant the responsiveness matters more than raw capability — a fast model you consult constantly is worth more than a slow one you avoid.

Keep reading

Running AI locally Ollama for business Local vs cloud, compared