ModelBrain

ModelBrain gives every AI assistant you use the same memory. It belongs to you, not to any one app.

The same memory works with every AI assistant you use, Claude, Cursor, and anything else that speaks MCP, so you're not locked into whichever app you happened to start in.

No account. Your vault stays on your computer by default; memories you share with an assistant are subject to that assistant's own data policies. You can read the memory file yourself, and delete anything in it whenever you want.

macOS and Windows builds are live now. Every page here still says plainly which parts aren't finished.

Monday

YouWe agreed to bill Kestrel Partners quarterly instead of monthly, at their finance team's request. Remember that.

AISaved. Kestrel Partners: quarterly billing, at their request.

Three weeks later · new chat, different app

YouWhy is Kestrel Partners on quarterly billing again?

AIYou agreed it on September 13, at their finance team's request, when they moved off monthly.

illustrative example · from your own note · Sep 13

Cognitive inference search

Point ModelBrain at a folder of notes on your computer, and every AI assistant you use can search it, then reason over what it finds.

Most memory tools find notes that look like your question: a ranked pile of snippets that resemble what you asked. Then your assistant has to sift through them and fill in the gaps itself.

ModelBrain works differently: a small AI model runs on your own computer, reads the notes the search turned up, and answers from what they actually say: one answer, with the specific notes it used cited, so you can check it yourself. It's built to tell you when the answer isn't in your notes, instead of guessing.

“What port does staging use?” → 5433, cited from infra/staging.md

“What's Dana's dentist's name?” → That isn't in your notes.

That's cognitive inference search: a local model that reads your notes and answers from them, with sources: not a keyword match, and not a guess.

Built for questions your notes can answer: facts, decisions, who's who, what was said. It isn't built for judgment calls yet.

The models live on your machine

Two AI models run entirely on your computer: one that finds candidate notes, one that reads them and answers from them. ModelBrain downloads and manages both: open-weight models, sized for your hardware, run locally via llama.cpp. No API key, no cloud inference, no per-question cost. The search and the reasoning happen on your machine; your assistant sees the question it asked and the answer it gets back, same as with any tool it uses.

ModelBrain picks the reasoning model automatically: a lighter model on modest hardware, a stronger one if you have the memory to spare. Either way, it runs without any setup. Models download once, the first time you turn Reasoning Mode on in the tray; after that, it works offline.

The reasoning model is read-only: it can read your notes to answer a question, but it can't write to or delete your memory. Before an answer goes back to your assistant, ModelBrain checks that the notes it cites exist and support what it says.

Every AI you use

The same memory, and the same reasoning, goes to whichever AI assistant you're in: Claude, Cursor, and other MCP-compatible assistants. The same folder becomes shared memory for all of them, not a feature locked to one app.

Setting it up, in full

1

Run one command

A single install command, nothing else to set up and no sign-up screen. It creates a vault on your computer and starts running quietly in the background.

2

Connect it to your AI

One entry in your assistant's settings and it's connected. Anything that supports MCP, Claude, Cursor and a growing list of others, works the same way.

3

Forget it's there

Carry on talking to your assistant: "remember this", "what did we decide on pricing?", "forget what I said about the venue". ModelBrain has no app of its own to open.

Free, and what "free" means here

Local memory is free, permanently: not a trial, not a limited plan, not a tier with an asterisk. Device sync and team plans are optional paid services, because they involve a computer ModelBrain runs, not one of yours; the storage they use is billed at a flat $0.05 per GB per month.

$0

No account, no card, no licence key, no expiry, and no cap on how much you remember. Commercial use included.

Unlimited memories

Recall across every session

Folder sync for your documents

Works with any backup tool you run

Works with any MCP assistant

Full export, any time

$6/mo Sync between your devices. The same vault on your laptop and your desktop, per device. Managed backup comes with it: encrypted on your machine before it's sent, so we hold a copy we can't read, and restores to a new laptop in one command. Storage is $0.05/GB/month. Coming soon, not built yet.

$6/node Shared team vault. Same per-node rate as device sync, plus SSO, an audit trail, and managed backup once it ships; backup and vault storage at $0.05/GB/month. Designed, not built yet.

Cancel and your vault keeps working; it never depended on us. Full pricing →

01

We tested whether it actually remembers, and published the results.

Any memory tool can list features. The question that matters is whether it brings back the right note, whether it points you at the note it actually used so you can check it, and whether it keeps doing that once you've got a thousand notes instead of ten.

So we built a test designed to catch our own mistakes, including the nasty cases, like two similar projects, a decision you changed your mind about, and a note you deleted. Across that test, it gave the right answer 86% of the time, and when it genuinely doesn't know something, it says so instead of guessing, every single time. It's still labeled beta, not a finished 1.0, and the full numbers are on the facts page.

The right note comes back
It shows you which note
It holds up as notes pile up
Deleted stays deleted

03

When you delete something, it's deleted, and you can see what changed.

"Forget that" does one of exactly two things, and both are written down. Ask it to forget a memory and it's gone from what gets recalled; it never quietly restores an older version behind the scenes. If you want an older version back, that's a separate, explicit "revert," never a side effect of forgetting. Ask it to forget a document you added, and the whole document goes, not a random paragraph of it.

Nothing you deleted comes back quietly. When you change your mind about something, the old version is kept as history you can look at on purpose, rather than competing with the new one behind your back.

A memory → deleted, not reverted
A document → all of it
Comes back on its own → never

Exact behavior →

One brain, every assistant, and if you want, your whole team

A memory that lives inside one app is a memory you lose when you switch apps, and one nobody else can benefit from. ModelBrain sits outside the assistant, so the same memories are there whichever one you open, and the same vault can be shared with the people you work with.

free · at release

Claude

Cursor

next year's

↓ one vault, read by all of them ↓

Your vault

Tell Claude something on Monday, ask Cursor about it on Thursday. Switch tools next year and the memory comes with you, because it was never inside the tool.

paid · designed

Dana

Marco

Priya

↓ private vaults + one shared ↓

The team's shared memory

"What did we agree with this client?" gets the same answer for everyone, and a new colleague's assistant knows the history on day one. Private memories are excluded from sharing before it's even considered.

Who it's for

People who use an AI assistant most days for real work and are tired of re-explaining the same context every morning. People who keep notes, decisions and client details they'd rather not hand to a company's servers. And developers who want the protocol details: those are on a page of their own, not spread across this one.

What it doesn't do yet

It's a beta on macOS and Windows. Linux isn't part of this launch.

It reads text documents: Word, PowerPoint, Excel, PDF, Markdown, plain text. Not scanned pages, which have no text in them to read.

It won't notice on its own that two of your notes contradict each other. You tell it which one is current.

Sharing memories between devices or with a team is designed, not built.

The vault isn't encrypted by ModelBrain today; turn on your computer's own disk encryption in the meantime. An encrypted-vault option is planned as a premium feature: why, in plain terms.