Design policies
A brain makes four decisions about each fact it extracts. Living Brain makes each one with a design policy, a score built from named signals that only counts once the item has passed the policy’s checks.
How a policy scores
Each policy raises a score for its signals and lowers it for its penalties. The checks run first. An item that fails one is excluded whatever it would have scored, so a page the caller isn’t allowed to see never becomes eligible because it is relevant.
The API doesn’t return policy scores, and how signals are weighted isn’t published. A score orders decisions inside the brain. It is not the probability that a fact is true.
Memory Worthiness Score
MWS decides how much maintenance a fact gets.
“Our enterprise plan caps seats at 50” will stay true for a while, is likely to be asked about and belongs to a project, so it lands in Durable and is maintained in a page. “Remind me to grab milk” scores low on durability and future use and stays Source-only. Provisional facts are kept and looked at again as more evidence arrives.
A low score deletes nothing. A source-only fact can still be found in its source. The mistake MWS has to avoid is a false exclusion, where a rare fact is skipped today and a question needs it next month.
Placement Fit Score
PFS decides which page a fact goes on.
PFS scores the fact against each candidate page and then measures the gap between the top two.
“Acme moved its launch to 15 June” and “Apex moved its launch to 15 July” embed almost identically, but they share no customer. The shared-entity and relationship signals keep the Acme fact off the Apex page. The outcome depends on the gap:
- Place: one page is clearly the best fit.
- Review: two pages fit almost equally well, so the fact is held rather than forced onto one.
- New location: no candidate fits. The brain starts a new page and keeps the source evidence.
A page that merges two customers reads smoothly and gives wrong answers about both, which is why a small gap goes to review.
Connection Resonance Score
CRS decides which pages are linked as worth reading together.
A page on “Postgres LISTEN/NOTIFY limits” and one on “Realtime notifications design” share little vocabulary, yet the first changes how you read the second. That is the test for a Strong link. Two pages that say the same thing lose score to the duplication penalty, and a link whose evidence has aged loses score to the staleness penalty.
Permission is checked before scoring, so a page is never linked to one the caller can’t see.
The connections the API returns today use a categorical rubric, strong, medium or weak, each
with a written reason. See Connection.
Re-entry Priority Score
RPS decides when a page comes back to you without being asked for. It scores the page against the current task and the current time.
“Send Acme the security review by Friday” scores high on urgency and open commitments, so it can trigger a Notification or appear in the Daily brief. A preference mentioned once months ago sits in Quiet context, where an agent can use it without interrupting anyone. Fatigue lowers the score of anything shown repeatedly and ignored.
Action-affecting means the page should shape what an agent does next. It does not give the agent permission to act.
How the four depend on each other
Each policy works on what the one before it produced. If extraction gets a fact wrong, placement puts the wrong fact on the right page. A connection only helps when retrieval follows it, and ranking can still drop a premise that earlier stages kept.
The scores measure different objects, so an MWS of 0.8 and an RPS of 0.8 can’t be compared, and no single number combines them. For the state of a whole brain, see Brain health.