Why I built AI Twin as a memory layer, not another assistant

I get asked, almost weekly, whether AI Twin is a competitor to ChatGPT or to Claude.

The answer is no.

Not because the product is too small to compete. Because competing with them is the wrong problem to solve.

The assistants are good. Getting better.

The thing they are bad at is the same thing they have been bad at since they launched.

They forget.

They forget what you told them last week. They forget what matters about your work. They forget the context that makes the difference between a generic answer and a useful one.

AI Twin is the memory layer. Built for individuals. Held privately. Retrievable from whichever assistant you reach for.

The choice to build it this way was deliberate. It came from three observations I think are still under-discussed in 2026.

The assistants are converging

A year ago, each of the big assistants had a clear strength.

Today the gap on most tasks is smaller than the variance within any single model's outputs.

I notice this in my own usage. I move between Claude, ChatGPT, and Gemini depending on the day. Sometimes the choice is rational. Sometimes I just open the one that loaded first.

If the assistants are converging, then the differentiation is no longer in which one you pick.

It is in what you bring with you.

Memory is the thing you bring.

People resist switching, and the resistance compounds

Once you have taught a tool about your work, your family, your context, switching feels expensive.

Not because the new tool is worse. Because you have to re-teach it everything.

The lock-in is not technical. It is emotional and informational.

You sit down with a new assistant. You realise you would have to explain who your co-founder is, what the supplier dispute was about, why the second project matters more than the first one, and you close the tab. Open the old one.

The way out is not "build a better assistant". The way out is build a memory that survives switching.

That is what AI Twin is.

The trust problem is the product problem

Most AI products in 2026 are not built to be trusted with sensitive personal information.

The terms of service usually disclaim. The training-data policies are unclear. The export options are weak.

People are using AI for personal life admin anyway. They are doing it nervously.

A product designed around trust from the first commit is a different product. Not the same product with a privacy page bolted on.

That distinction matters more than I expected it would.

What this means in practice

The design constraints follow from those three observations.

Smaller product surface. We do not need to win the assistant war.

Slower trust building. We publish what we are certified to, not what we hope to be.

Higher friction on some flows. Consent is real, which means clicks, which means slower onboarding.

Lower expectations for "wow" moments. Memory layers prepare. They do not perform.

The trade-off accepted: less surface area, more trust.

What I would have built if I were not me

There is a version of AI Twin that would have been faster to build and easier to demo.

A chatbot with a calming voice. A productivity app with AI sparkles. Something that fits the existing categories.

We did not build that one.

The category we are in is smaller, quieter, and harder to explain in a thirty-second video.

It is also the one I want to use myself. And the one I think will look obvious five years from now.

The next round of AI products will be judged on what they hold of us, not what they answer for us.

The memory layer is where that judgment lands.

More soon.