One brain on your card, shared by the whole stack
Every local-first tool wants its own model on your GPU, and a single card runs out of room to hold them all. Arbiter keeps one resident brain and lets the rest of the stack borrow it, so Curator's memory work and Arbiter's own routing run on a single loaded model instead of a pile of duplicates.
The promise of local-first AI is that the model runs on your machine instead of someone’s cloud. The catch nobody mentions is what happens when you own more than one local-first tool. Each one wants its own model resident on your GPU. Two of them already crowd the card. Five will not fit on a typical card, and the ones that load spend their first seconds waking up before they can answer. The local-first dream turns into a fight over the same slab of memory.
Arbiter takes the other road. It keeps one brain resident on your GPU and lets the rest of the stack borrow it.
One model, many callers
Arbiter already runs a local model for its own reasoning: the routing, the triage, the operations chat. That model sits loaded on your card whether or not anything is asking it a question this second. So the wasteful thing would be for the app next door to load a second one just like it. Instead, Curator sends its memory work to Arbiter’s brain rather than standing up its own. One resident model answers for both. The design goes further than the one live example: the stack is built so an app that needs coordination-class inference can lean on the same brain instead of shipping yet another model onto the same GPU.
You buy the card once
The economics are the whole point. A local model is only free of a cloud bill. It is not free of the hardware it sits on, and GPU memory is the scarce thing in this house. When every tool insists on its own model, you either buy a bigger card or you start choosing which tools get to run. When they share one resident brain, that pressure goes away. You pay for the card once, and the stack lives inside that budget instead of multiplying it. We are keeping specific memory and latency numbers off this page on purpose, because they depend entirely on your card, but the shape does not change: one model on the GPU is cheaper to house than five.
It is already warm
A model that is already loaded does not make you wait for it to wake up. Because Arbiter’s brain stays resident, the app that borrows it inherits a warm model instead of paying a cold start of its own. The small, constant work a stack generates all day, classifying a note, summarizing a run, deciding where something should go, lands on something that is already ready. Shared does not mean slower here. It means nobody else has to boot their own copy first.
Still on your side of the line
Sharing the brain does not widen who can see your work. The one resident model lives on your GPU, and the inference the stack borrows never leaves the box to get answered. This is the same principle the whole sovereign stack is built on, applied to its own reasoning: your compute, your keys, your decisions, on hardware you own. We wrote about why Arbiter’s brain runs on your GPU in the first place, and about how that brain recently gained more room. Sharing it is what turns one good local model into the reasoning layer for a whole stack.
To be honest about the boundary: the shared brain is for coordination-class work, not for being a coding agent. Heavy reasoning still routes out to whatever frontier provider you have chosen for it, under the cost and safety policy you set. What the shared brain removes is the other thing, the pile of duplicate local models each app would otherwise load just to handle its own small decisions.
The conductor holds the brain
It fits the rest of Arbiter’s job. The layer that owns the terminal, supervises the fleet, and dispatches Warden, TheAuditor, and BenchProctor is also the natural place to keep the one brain they can all reach. The conductor holds the baton, and now it holds the brain the whole orchestra shares.
Arbiter is a proprietary working alpha in public binary release preparation. Follow the public release channel or RSS feed for updates.