Memory is not storage. It is a neuron.
Most AI memory is a database bolted onto a frozen model. We think remembering and thinking should be the same act.
Today's models are brilliant and they are also static. They answer from a mind that stopped changing the day training ended. When memory exists at all, it is a retrieval system stapled to the side: search, fetch, paste. It is recall without growth.
The problem with retrieval
Retrieval treats the past as a filing cabinet. You write something down, and later you go look for it. The model never changes — it just gets handed a few paragraphs and asked to pretend it remembered them.
That works for trivia. It fails for anything that has to accumulate: a companion that should know you better next month, a robot that should not relearn the same building every morning, an agent that should stop repeating yesterday's mistakes.
Memory inside the network
We take a different position. Memory should live inside the network, the way it lives inside a brain — as structure that is changed by experience, not as rows in a table beside it.
When memory is a neuron rather than a record:
- Remembering and thinking become the same act.
- Context has place and relation, not just text.
- The system grows from what it sees instead of resetting to zero.
This is harder. It is also the only version of memory that lets a machine feel less like a tool and more like a mind.