Research notes

The AI stack problem

22 Sep ’26
appmodelsomeone else's cloudyou, outsideappmodel · swappablethe substrateyou, at the base

Every layer of today’s AI is built to forget you.

Look at how intelligence reaches you today.

A model, rented by the token. An assistant that wraps it and keeps your history in its own account. Apps stacked on top, each holding its own version of who you are. Nothing beneath them keeps your memory, your language and your way of thinking in one place that you own.

So every layer starts from zero. Every layer asks again. And when the model changes, everything you taught the last one is gone.

The result is intelligence that gets more expensive the more you use it, and never quite gets to know you.

Our answer is not another layer on top. It is the layer underneath: a substrate that keeps what you teach it, stays on a device you hold, and lets any model plug in and become yours. Sharper and cheaper the longer it stays with you.

That is what we are researching now, in the open. Production is the goal.

ritam — from ṛta (ऋत), the silent, rhythmic order of truth.