A model with unlimited context that remembers what you told it
Every new session starts with you saying it all again: who you are, what you are working on, what you told it last week. Carpathian models trained in Veritate remember. What you tell them becomes part of the model itself, so nothing has to be pasted back in and nothing piles up in front of your question. We do not run these in the chat on this site. They are built to run on your own computer.
We taught it, then we let it sleep
Fifty things it had never heard, each told once. Three evenings later it answered ninety-four questions out of a hundred about them from memory, with nothing put in front of it.
- DayYou talk to itWhatever you tell it during the day is kept as the day's experience.
- EveningIt sleepsWhile nobody is using the machine, it goes back over what happened.
- AlongsideIt revisits older daysEarlier conversations come back through with the new ones, so what it already knew stays.
- GentlyIt learns in small stepsA new fact settles in rather than landing on top of something else.
- MorningIt knowsYou start a new session and it is already there. Nothing to paste in, nothing to look up.
Why call it sleep?
Brains solved this a long time ago, and the outline of the solution is well documented. You do not burn new memories straight into cortex while you are using it, because that would overwrite what is already there.
A fast temporary store catches the day's experience. During sleep the brain replays that experience into the slow permanent store, gently, mixed with older memories, at an intensity that does not bulldoze anything.
A short training run during idle time, dosed to how much happened, mixing new material with a quarter of old, at a learning rate low enough that the model bends rather than snaps. We call each run a night.
What to expect
How long it takes
After the first night you barely see a difference, a couple of nights in, most of what it learned has become part of the neural network.
When it stops
The model checks itself against the way it used to answer, once it has changed enough for one week, it stops.
Asking it backwards
Tell it Sarah lives in Boulder, and then ask who lives in Boulder, it will remember, Sarah.
It still talks the way it did before
Teaching a model new facts usually costs it something, it picks up what you taught it and loses the thread of a conversation.
Ours comes through the nights sounding the same, it stays in character, it finishes what it starts, and it repeats itself slightly less than before.
One of our checks asks the model what it just said, in a fresh request with nothing in front of it to read, before the nights it always said it did not know.
After them it answers correctly about half the time, and when it cannot, it says so instead of making something up.
Where this goes
The software already runs on the same machine that answers you, it waits for the quiet hours and does the night's work then. Nothing leaves the machine, nothing is copied out, nothing is pasted back in the morning.
What that adds up to is a model that grows into the person using it. Talk to it today and tomorrow it knows, talk to it for a year and it knows you well enough that starting over somewhere else would feel like losing something.
Every machine that can run a model can run this, that is the whole point. The memory is yours, it sits on your desk, and it goes where you go.
Go tell it something
The model in our public chat is the one that sleeps, it is the only one we run this on. Tell it something today, come back tomorrow, and see what it kept.