Docs / Hardware requirements & model tiers
Remembering and plain search run on essentially any machine. Reasoning — the part that connects facts across notes and checks its own answer — runs a real local model, so it scales itself to what your computer can actually carry. You never pick a model by name; ModelBrain measures your RAM and CPU cores and picks the tier for you.
RAM here means free RAM available to ModelBrain, not your machine's total installed RAM. Cores are your CPU's performance cores, not a raw thread count.
Only the model. Every reasoning feature — looking up more context, verifying its own citations, retrying a weak answer — runs identically on every tier; nothing is switched off to make a slower machine fit. A machine that can't run even the lightest model in time says so immediately (hardware_too_slow) instead of quietly returning a worse answer.
The off tier changes nothing about search, remembering, or recall — only the reasoning pass that connects facts across notes is unavailable.
Auto-selection is the default and usually the right call. If you want to override it — trade answer quality for speed on a borderline machine, or force a lighter tier for a faster response — the reasoning_tier setting takes auto, minimal, default, high, or off. Run diagnostics at any time to see which tier and model are actually active on this machine right now.
Plain search runs on essentially anything. Reasoning needs at least 7.5 GB of free RAM and 4 CPU cores for the lightest tier; 15 GB and 8 cores for the default tier; 30 GB and 8 cores for the top tier.
Reasoning reports hardware_too_slow immediately rather than running a crippled version of itself. Plain search, remembering, and recall all still work fully.
Yes, via the reasoning_tier setting (auto, minimal, default, high, or off). diagnostics shows which tier and model are actually active.