# The Brain: self-healing documentation

> Two layers: the App Brain knows your product and self-heals; a private Customer Brain remembers each customer. So replies are correct and personal.

Canonical: https://peeve.ai/product/brain/

The Brain is Peeve's live model of your product's interface: every element stored by role, accessible name, landmark position and neighbours. Every session is a measurement; a confidence that falls flags the route for re-capture. A rename merges automatically; a structural change waits ~30 seconds for your confirmation. Chat widgets can't do this. They never touch your interface, so they have no signal to collect.

## Two layers: App Brain + Customer Brain
Peeve grounds every answer in the App Brain (a self-healing model of your product, shared across all your customers) and, on top of it, keeps a private Customer Brain for each individual customer: a memory of who they are, their recurring issues, their preferences, and how their past issues were resolved. The Customer Brain is distilled automatically from that customer's own conversations (no CRM tagging) and grows every time they interact, so a reply is correct from the App Brain and personal from the Customer Brain. It personalizes email and social (WhatsApp, Messenger, Instagram) replies today, with the in-app widget as a fast-follow. Each customer's memory is encrypted at rest, isolated to your workspace, never shown publicly, and never used to train outside AI models.


## More questions

### How does the Peeve Brain work?

The Brain is Peeve's live model of your product's interface, rebuilt from observation every session. When a route's confidence falls, Peeve flags it, re-captures it on the next visit, and diffs it against the map. A rename merges automatically and a structural change waits about 30 seconds for your confirmation.

### Does Peeve hallucinate or give wrong answers?

Peeve answers from a live model of your product with the source attached, carries a confidence score on every element it finds, and below 0.60 it stops and hands off rather than guessing. Because knowledge comes from watching the live product, it does not invent a button or a policy that does not exist.

### Does AI customer support make things up (hallucinate)?

It can when it is ungrounded or forced to answer. The safeguard is grounding plus permission to say I do not know. Peeve answers from a live model of your product with the source attached, carries a confidence score on every element it finds, and below 0.60 it stops and hands off rather than guessing, so it does not invent a policy or a button that does not exist.

### How do you stop an AI support agent from giving wrong answers?

You ground it in real, current knowledge, gate it on confidence, and let it escalate. Peeve rebuilds its model of your interface from observation every session, scores how confidently each element was found, and hands off below its threshold. Because knowledge comes from watching the live product, a change in the product is a change in the answer.

### How does AI customer support stay accurate when my product keeps changing?

Most tools drift because they train on a snapshot and never update. Peeve treats every session as a measurement: when a route's confidence falls, it flags the route, re-captures it on the next visit, and diffs it against the map. A rename merges automatically; a structural change waits about 30 seconds for your confirmation.

### How do you keep AI support answers from going out of date?

Continuous grounding beats a one-time training dump. Peeve does not rely on a static export of your help center; it re-reads your live interface as users move through it, so the knowledge tracks the product in real time instead of decaying between manual updates.

### What happens when the AI does not know the answer?

The honest behavior is to say so and hand off, not to guess. Peeve stops when its confidence is below threshold, opens a ticket in your helpdesk with the intent in the user's words and every step attempted, and conversations where it found nothing and said so are never billed.

### Can I see why the AI gave a particular answer?

Traceability matters for trust. Peeve attaches the source to answers, records the route and confidence it used on every session, and writes a signed audit entry for every backend action, so you can see what it said, why, and what it did.

### How can AI resolve issues if I do not have a big knowledge base?

A large help center is not required. Peeve learns from the product itself by observing the live interface, and can migrate what help content you do have from tools like Intercom, Chatbase, Tidio, Crisp, Freshchat, Help Scout or Ada. It improves continuously as it is used rather than needing everything written first.