It already knows the project.
Decisions, rules, sources and earlier Missions, compiled for every goal with a reason per item. Nothing has to be said twice.
“Lead with speed, not price” came from memory, not from the person. The story was written that way, the page was changed that way, and nobody repeated it.
✓ confirmed 12 Oct · used by 3 runs · IllustrationEverything it may read, in one place.
One Brain per project. Connected with your own token, stored encrypted and shown as its last four. A connector document re-syncs on a schedule you turn on, after asking the source whether anything changed.
- Repositorynorthstar/web · release/v4chunked by declaration · symbols · import edges · git history
- Filesbrand/guide.md · 14 interviewsdropped on the composer or uploaded · indexed inline
- NotionLaunch v4 · decisionspages, with a pasted token
- LinearNS-412 · launch targetissues, with a pasted token
- Slack#launchchannels, pasted token or OAuth
- Gmailsupplier threadsthreads, through OAuth
- Drivecustomer researchfiles, through OAuth
- Airtableordersbases, with a pasted token
It remembers reasons, not transcripts.
Decisions, rules, rejected approaches and bugs, each with a source and a state. A memory whose files changed after it was confirmed is doubted until checked. When a memory contradicts the code, the code wins.
Every goal gets a plan.
The goal is read for what it names. Paths, identifiers and error lines become weighted lanes; what each lane found is ranked, packed to a budget, and shown on the Context card with its reason.
- path×2the goal names a file: PricingHero.tsx
- symbol×1.6the goal names an identifier: PricingHero
- vector×1.2UI goals describe what is seen, not how it is named
- graph×1.2neighbours of the best files
- lexical×1the goal’s own words
- memory×1decisions and rules the project already made
- critique×1the last design critiques say what already looks wrong
Two memories disagree. You keep one.
Pairs that cannot both be true are surfaced on the Context card: the same measure with different numbers, or opposite polarity. Boomship says which outranks which. A person decides.
Answers only from sources.
A question is searched like the Brain; what it finds is numbered. The model sees only those sources and must cite them. Citations are checked afterwards. Nothing matched, no model call.
Because positioning was decided that way and confirmed twice[1]: eleven of fourteen interviewed teams chose their supplier for setup speed[2]. The hero copy in [3] was changed to say so in Mission 0218.
- [1]memoryLaunch positioning leads with speed, not price.confirmed decision · from a person
- [2]documentCustomer interviews · Driveconnector · synced 2 h ago
- [3]codecomponents/PricingHero.tsx · 12–15path lane · pinned
One graph, lit where the goal walks.
Earlier Missions are evidence too.
Runs on similar goals are cited from their receipts and what happened after: held up, undone, disproven. Never from a model’s summary of them.
Your model may change. Your company context shouldn’t.
Claude, GPT, Gemini or OpenRouter do the thinking. The Brain, its files, connections and memories belong to the project, and go with it to the next model.
- Decision. Launch positioning leads with speed, not price.
- Customer. Teams love the setup but struggle with the first configuration.
- Brand. No gradients in editorial materials.
- Calendar. Launch target · Nov 18.
- Product. v4 candidate ready. 1,214 symbols indexed.