Answers
How does an AI agency keep one client's rules out of another client's work?
By keeping each client in its own project, and the agency's own way of working in a separate one. Memory in Arroway is scoped to a project and never global: an assistant working on an account reads that client's sanctioned rules plus the agency's house method, and nothing from any other client — because the other accounts are not in scope for that read, not because someone remembered to leave them out. The same person and the same assistant can work six accounts in a day without the sixth inheriting the voice of the first.
Last updated September 3, 2026
Two kinds of rule, and only one of them should travel
An agency accumulates two sets of standing rules, and they behave in opposite ways. The client's set is theirs — voice, terminology, what they refuse, who signs off — and it must not surface anywhere near another account. The agency's set is a large part of why clients hire you — how you scope, how you review, what you never ship — and it should surface on every account. Keeping both in one place is what produces the failure everyone in this business recognises: a deliverable that reads like a different client's.
One project per client, and the read is scoped to it
Each account is a project with its own members and its own material. An assistant asks for the project it is working in and receives that project's rules in force; there is no global set riding along quietly, and no per-account preamble that someone has to keep in sync. Taking someone off an account ends their access to it, and what they wrote there stays with the account under their name.
The house method is the one thing every account shares
The agency's own standards live in their own project and are read alongside the client's, which is how an assistant follows your review process on an account that has never heard of it. That set is also where scale actually comes from: change how the agency scopes a build once, and every account is working the new way on its next read — instead of the change reaching the two accounts whose preamble somebody remembered to edit.
What it looks like in practice
An agency runs eleven accounts with four people and the same assistants. Two of the clients are in the same category. One of them will not use a particular claim, on legal advice; the other has a house style that leans on exactly that kind of claim. Both rules were mentioned on calls and then lived inside a long preamble that whoever was on the account pasted at the start of the day. One morning the preamble for the wrong account got pasted. The draft came back in the other client's voice, carrying the claim they refuse. Nobody was careless — the two rule sets were in one place, and all that separated them was a person picking the right block of text under time pressure. Scoped to a project, that is not a mistake available to anyone: the wrong set is not in the read.
Questions people ask about this
- Someone works on two accounts. Can they see both?
- Through their membership of each project, and one project at a time — the assistant's read is scoped to the account it is working in, not to everything the person can reach. Access is granted and ended per client, which is also what makes the question answerable later: you can say who could see an account and when, instead of who could see everything.
- How do we show a client their rules were followed?
- Every rule in force carries its provenance: what was decided, when, and who approved it. So the answer to a client who asks is the record itself, in the state the assistant read it, rather than a reconstruction written afterwards. What the record does not do is prove that a particular deliverable came out of it — that is what your own review is for.
- What happens when an engagement ends?
- The project stays with the account, and the people on it lose access when they are removed, while what they wrote stays with its author and its history intact. If the client returns a year later, the decisions that governed the last engagement are still there and still say who took them. If the relationship ends for good, the material can be exported and the members removed. Neither outcome depends on somebody remembering to tidy up in the moment.
Where this is verifiable
Product documentation on this site (How it works, Install), the project and membership rules in the sanctioned product spec, and the Privacy Policy for what happens to a person's record when they leave a project. Everything described here is behaviour the tools apply today, not roadmap.
https://www.arroway.app/en/answers/shared-memory-for-ai-agencies