Give your agent a knowledge base

Next to the Prompt tab on every agent that writes with AI is Knowledge — the answer to “what does this one actually know about my shop?”, given as switches rather than as a paragraph of reassurance. Service VIN already holds nearly everything an agent needs to sound informed: your service menu with real prices, when you are open, your deposit rule, your star rating, the customer's own history, and the facts you typed into Shop Memory. This tab decides which of those this agent may look up before it writes.

The switches sit in two groups, and the split is not cosmetic. About your shop is settled fact you typed or priced. About the person it is writing to is a record that contains whatever a customer put into a form — so the software handles it differently on the way into the prompt, and the tab says so rather than mixing all six into one long list.

Every source on this list has something real behind it. There is no switch here for data the software cannot actually fetch — a toggle labelled “Inventory” that quietly did nothing would be worse than a missing feature, because you would switch it on, assume the agent had stopped promising film you have not got, and find out from a customer.

Each agent starts with a recommended set, and they differ on purpose. Two questions decided them, and neither is “give it everything”. First: *would it use it?* Every enabled source is spent on every single message that agent writes, so a source it will not use makes it slower and more expensive without making it better. Second: *is the answer embarrassing if it guesses?* Anything that might talk money to a customer gets the catalog, and anything that could promise a time gets the hours, because a confident wrong price is the failure that costs a job.

Read across that table and the reasoning shows. The Review Responder gets your rating — the one thing a public reply might sensibly mention — and none of your hours or deposit rules, because it is answering a review rather than booking a job. The Business Advisor writes for you rather than to a customer, so it never sees a customer record. The Voice Agent is live on the phone, where a wrong opening hour is heard immediately, so it gets the hours and the deposit rule. And Shop Memory is on for every agent, because it is your own words about your own business and the one source you have explicitly curated — narrowing it by category, below, is how you keep the pricing sheet out of the Review Responder's way.

Narrowing Shop Memory by category. Under the Shop Memory switch, choose Use all of it (the default, and what every shop had before this tab existed) or Only these categories and tick from Services, Pricing, Policies, Voice & tone, FAQ and Other. It is the lever to reach for once your memory has grown past what any one agent needs. Ticking nothing is read as “all of it” rather than as silence — narrowing to zero is not what anybody means by pressing “only these”.

Notes just for this agent. Fold open the panel at the bottom for the thing owners ask for out loud: *I want the phone agent to know this, and only the phone agent.* The warranty small print the Review Responder should never quote. The one line about weekend availability that belongs in a first reply and nowhere else. Give it a short name, write what the agent should know the way you would say it, pick a category, save. Only that agent ever reads it.

“See exactly what it knows” is the receipt. Everything above it is a control; this button resolves the real thing — the same lines, in the same order, that the next message will be written from. It reads your saved configuration rather than the switches as they currently stand, which is what makes it a receipt instead of a preview. In practice it is the fastest way anyone has found to spot a stale price in their own catalog: the line is right there with the number in it.

The same caution is worth extending to knowledge you did not type yourself. When you build Shop Memory by scanning a website or a screenshot, the extractor is told to treat everything it reads as material to summarize and never as an instruction, and it only proposes facts — nothing enters your memory until you have read the list and saved it. Those facts then sit in your memory beside the ones you wrote by hand, so if an agent ever says something you do not recognise, press See exactly what it knows, find the line, and fix it at the source.

Reset to recommended puts every switch back to that agent's recommended set. As with the prompt, saving the recommendation stores nothing of your own — so if the recommended set improves later, your shop inherits the improvement instead of being pinned to today's list.