I use AI every day.
Probably far more than most people.
And the more I've used it, the more I've become convinced that we are concentrating on the wrong thing.
We keep looking at the model.
Everyone is talking about which AI model is best.
OpenAI launches something new. Google responds. Anthropic improves Claude. Another model appears. Benchmarks move. People switch backwards and forwards.
But the model isn't actually the thing I'm becoming attached to.
It's the context.
The conversations. The projects. The decisions. The people I've mentioned. What I'm trying to achieve. Why I made a particular choice six months ago. The things I said I'd come back to. The little pieces of information which, over time, mean an AI doesn't have to start from scratch every time I speak to it.
That's where the real value starts to build.
And that creates a question I don't think we've really answered yet:
Who should that intelligence belong to?
What happens when the model changes?
Think about how most of us use AI today.
You open ChatGPT, Claude, Gemini or whichever model you prefer and start talking.
Over time it may learn things about you. Some products now have memory. Others can connect to your email, files or calendar.
That's useful.
But you're effectively building that relationship inside someone else's product.
If you change provider, what happens?
What if the best model next year comes from a company that barely exists today?
Do you start again?
Do you have to explain who you are, what you're working on and hundreds of decisions you've already made?
And what happens in ten years?
I don't think the answer can be that we each build increasingly valuable digital context inside whichever AI company happens to have the best model at the time.
Models will change.
You shouldn't have to.
Memory is only part of it.
This was one of the things I struggled to put my finger on when we started building Arit.
At first it was tempting to describe the problem as AI memory.
But I don't think memory goes far enough.
Remembering that my son's name is Sam is memory.
Knowing that Sam runs a business, understanding which projects I've discussed with him, remembering a decision we made three months ago, knowing that something is unresolved and bringing it back to me at the right moment is something more useful.
And even that isn't enough if the information can't be trusted.
An AI can infer something about me. That doesn't necessarily make it true.
I can brainstorm an idea with an AI. That doesn't mean I've decided to do it.
An AI can suggest an action. That doesn't mean I've authorised it.
Those distinctions become much more important as AI moves beyond answering questions and starts doing things.
Personal Intelligence, not a digital clone.
That's why I've come to think of this as Personal Intelligence rather than simply memory.
Not an AI pretending to be me.
Not a digital clone.
And not one enormous database containing everything I've ever said.
I mean an intelligence layer belonging to the individual which can preserve useful context, knowledge, history and continuity over time — while still knowing the difference between something discussed, something inferred, something remembered and something actually authorised.
This is probably the simplest way I can explain the idea.
Imagine you've spent five years using an AI.
It understands your work. Your projects. Important relationships. Decisions you've made. Things you are waiting for. Documents you've worked on. What matters to you.
Then a substantially better AI model appears.
You should be able to use it.
Without losing the previous five years.
In the way I see it, the model should eventually become more like an engine.
You might prefer one today and another tomorrow.
Your Personal Intelligence sits above that.
It decides what information is relevant to a particular request and gives the model enough context to help you.
The model doesn't need your entire life every time you ask a question.
And your accumulated intelligence shouldn't disappear because you changed engine.
That principle became central to how we're building Arit.
Helping is not authority.
The more an AI knows about you, the more useful it can become.
But the more it knows, the more important control becomes.
Suppose I say:
“I really need to email Sam about that tomorrow.”
An intelligent system might quite reasonably recognise an intention.
But that's not permission to send Sam an email.
Or perhaps I ask:
“What would you say to him?”
Again, that's not permission to send it.
This sounds obvious when written down.
I'm not sure it will remain obvious as AI agents get better at taking actions.
We've therefore adopted some very simple principles while building Arit:
- Conversation is not authority.
- Inference is not truth.
- Capability is not permission.
- A plan is not permission.
These might sound like technical design rules.
They're actually human ones.
The system should understand the difference between helping me think and acting on my behalf.
Does knowing you mean sharing everything?
No.
This is another assumption worth challenging.
If I ask an AI to help write an email to somebody, does the AI provider need access to every conversation I've ever had with Arit?
Of course not.
It needs the relevant information for that job.
If I ask about a project, it may need context from that project.
If I'm asking a general question, it may need almost nothing personal at all.
A Personal Intelligence system should therefore be capable of knowing a great deal about its owner while being selective about what it shares with the AI being used to reason about a particular question.
That difference is going to matter.
Because in my view, the answer to increasingly capable AI cannot simply be:
Give the model everything and hope for the best.
More than an assistant.
I think the term AI assistant will eventually feel too small.
An assistant normally helps with something happening now.
What I'm describing is continuity.
Imagine being able to ask:
“What happened with that architect we spoke to last year?”
“Before I meet Sarah tomorrow, remind me of the important things we've discussed.”
“Where did I leave that idea about the new product?”
Or simply: “What am I forgetting?”
For those questions to become genuinely useful, the system needs more than access to a clever language model.
It needs continuity.
It needs structure.
It needs to understand what belongs to you.
And increasingly, it needs governance around what it can do with it.
The thread should stay with you.
This thinking is what led us to Arit.
We're still building it, and I don't want to pretend we've solved every part of this problem.
We haven't.
But the direction is clear.
Arit is being built as a persistent, governed Personal Intelligence system.
The ambition is that your useful context stays with you, rather than becoming inseparable from whichever AI model you happen to be using.
Models can improve.
Providers can change.
New interfaces will appear.
AI will become embedded in phones, cars, watches and things we haven't thought of yet.
But the thread of your life and work shouldn't have to start again each time the technology changes.
That's the part I think we need to get right.
Because perhaps the most valuable thing we will build with AI over the next decade won't be another model.
It will be the intelligence each of us builds around ourselves.
And if that happens, I believe it should belong to us.