Glossary
Belief propagation
You changed your mind about one thing. The cost is never that one thing. It is the six conclusions still standing on it, which all still look correct.
Belief propagation (operation). When a held belief changes, everything resting on it is located and re-examined rather than left in place. Implemented in Atlas, an open-source local-first system that satisfies the full AGM postulate set for rational belief revision.
Impact radius
Take an ordinary example. You decide your best customer is not the segment you have been targeting for three years. One belief, changed, and you would say you have updated.
Now count what was built on it.
- Who your marketing speaks to
- Which objections your sales process is designed to handle
- What onboarding assumes about the person arriving
- Which features got prioritised and which got shelved
- What counts as a good week in customer success
- The standard you apply when deciding whether a deal is worth taking
None of those are marked as dependants of the belief you changed. Each was arrived at soundly, from a base that was correct at the time. Each still reads as correct now, because its internal reasoning is intact.
Only its base moved.
That is why the naive picture of updating is wrong. Changing your mind is not editing one record. It is a shockwave through a structure whose shape nobody wrote down, and the size of the affected region is not visible from the change itself.
Why memory alone cannot do this
A memory system has two operations. Store something, recall something. Both useful, neither sufficient here.
Adding a newer note is still storing. Both versions now sit in the pile, nothing marks one as superseded, and conflicts get resolved on recency, which is a proxy for authority rather than authority itself. The most recent thing you said about a subject is frequently the least considered thing you said about it.
Revision is a third operation and it is qualitatively different: this position replaces that one, and every conclusion resting on the replaced position is now suspect. That is not a bigger version of storing. It requires knowing what depends on what, which memory does not track because it was never built to.
A web understands what depends on them.
The part everyone gets wrong in the other direction
Having seen the problem, the tempting fix is to re-open everything whenever anything changes. Full recomputation. It is thorough, it is easy to reason about, and it is unusable.
Most of what rested on the old belief still holds. It holds for reasons that had nothing to do with the belief that moved, or it was overdetermined, or the belief was one of several supports. A system that flags all of it hands you a list you cannot act on, and the third time that happens you stop reading the list, which returns you to where you started with extra steps.
So the useful behaviour has three parts.
- Preserve what still works. Most of it.
- Update the connections that are clear. Where the dependency is direct and the new position implies the new answer, make the change.
- Return the true exceptions. The cases where the change genuinely reopens a judgment call go back to the person, and there are far fewer of them than the full dependency list.
That third pile is the only thing anyone should be asked to look at, and keeping it small is the actual engineering problem.
Why this decides whether captured judgment is worth having
Judgment that cannot update is a liability with a good reputation.
A captured layer that never revises becomes an accurate model of the person you were when it was built. Every month it stays confident and gets a little less correct, and because the failure is silent there is no moment where it announces that it has drifted. It simply keeps making the calls you would have made two years ago, which you would recognise as wrong instantly if you were looking, and the entire reason it exists is so that you are not.
Which is why revision is not a feature sitting alongside capture. It is the condition under which capture is worth doing at all.
The formal grounding, and why it is named
Belief revision has been studied for decades. The standard framework, AGM, specifies postulates a rational revision operation must satisfy when new information conflicts with existing commitments: what has to be preserved, what must give way, and what may not be arbitrary about the choice.
What has been rare is any consumer-facing system implementing them. Atlas is open source and implements the full set, which is the difference between claiming that a system updates properly and being able to show the check.
Frequently asked
What is belief propagation in this context?
The operation that runs when a captured belief changes: everything that was resting on the old belief is located and re-examined, rather than left in place. Without it, a system keeps serving confident conclusions derived from a position you have already abandoned.
Why is this harder than just updating the fact?
Because the fact is one entry and the conclusions built on it are many, they are not labelled as dependants, and each one still reads as correct on its own. Nothing in a stale conclusion looks stale. It was sound when it was formed and its reasoning is intact. Only its base moved.
Is belief revision a real field or a marketing term?
A formal field. The standard framework is known as AGM, after Alchourrón, Gärdenfors and Makinson, and it sets out postulates that a rational revision operation has to satisfy when new information conflicts with what is already held. Atlas implements the full postulate set and is open source, so the claim can be checked rather than believed.
Should the system change everything downstream automatically?
No. Blind global revision is its own failure. Most of what rested on the old belief still holds for reasons that had nothing to do with it. The correct behaviour is to preserve what still stands, update the links that are clear, and bring the genuine exceptions back to you, which is a much smaller pile than the full dependency list.
What does this failure look like in practice?
It looks like good work. That is the whole difficulty. The system produces well-reasoned output based on a position you dropped six months ago, with no flag, no uncertainty and no way to spot it except by already knowing. The failure is silent, which is exactly what makes it expensive.
Where this sits
What this operation maintains is described at Imprint, and the category is defined at What is Imprinted AI. Atlas was published open and free because captured judgment that cannot change with you eventually becomes a museum of the person you used to be.
Related terms
- Imprint. The set this operation maintains.
- Digital DNA. The wider model that also has to stay current.
- Imprinted AI vs AI memory. Why store and recall alone cannot do this.
- An Imprint vs an SOP. The same staleness, in a document you already own.
Where the term comes from
This is one entry in the vocabulary of a longer argument. The full glossary has fifteen terms. The case they belong to runs about 23,000 words, it is free, and there is no email gate on it.
Read The A.I. Business Manifesto
Nothing on this page is for sale. Quote it, argue with it, or pass it on.
Last updated: 28 July 2026