> ## Documentation Index
> Fetch the complete documentation index at: https://project-brain.rocheaiops.net/llms.txt
> Use this file to discover all available pages before exploring further.

# How B.R.A.I.N. thinks

> A plain-language account of retrieval, evidence, provider reasoning, response, and persistence.

B.R.A.I.N. does not treat a model response as memory. It receives an input, gathers applicable durable records and document knowledge, preserves the evidence behind them, asks a provider for bounded reasoning when needed, and records only the durable event and accepted learning.

<a href="/assets/visuals/brain-system-context.svg" target="_blank" rel="noreferrer" aria-label="Open full-size diagram: Cognitive system context"><img src="https://mintcdn.com/roche-ai-ops/zV65QLx2v4-e7KNO/assets/visuals/brain-system-context.svg?fit=max&auto=format&n=zV65QLx2v4-e7KNO&q=85&s=a009613b7ea420e14818da8962b09e93" alt="Cognitive system context" noZoom width="1440" height="820" data-path="assets/visuals/brain-system-context.svg" /><span className="sr-only">Open full-size diagram: Cognitive system context</span></a>

A response can be useful without becoming a belief. A document can be relevant without becoming an owner preference. Authority answers *who is entitled to establish this?* Confidence answers *how certain is this interpretation?* They are independent fields.

Example: an owner says “Use Hostinger.” That can become an active record with owner authority. A model infers a pattern from prior sessions; that remains a candidate unless attributed validation promotes it. See [trust and authority](/concepts/trust-and-authority) and [canonical state](/architecture/canonical-state).
