Twelve clients, no mixed-up memory

Agencies and freelancersDev agency with many client repos

Our agency works on 12 client codebases. We wanted each to have its own memory, and never one client's decisions showing up in another's work.

How do I keep AI agent memory separate for each client repo at my agency?

Run arcroom init --project <id> in each client repo so each one is its own ArcRoom project. Recall and find_task only look inside that project, so one client's decisions never show up in another's work, and each folder brings its own memory as developers move between clients. Your agency's shared standards have to be repeated in each project.

Also asked as: How do I stop one client's decisions leaking into another client's work with AI agents? · How do I set up separate agent memory per codebase? · Can AI agents keep different memory for different client projects?

The work

Each repo is its own ArcRoom project (arcroom init --project <id>).

ArcRoom keeps
  • Reused answer
  • Recall
The next agent gets

[client-b repo] Team decisions (client-b): Next.js pages router only, client won't upgrade.

Step by step

  1. Each repo is its own ArcRoom project (arcroom init --project <id>).
  2. Recall and find_task only look inside that project.
  3. Developers move between clients and each folder brings its own memory.

What the agent sees

[client-b repo] Team decisions (client-b): Next.js pages router only, client won't upgrade.

Compared with today's tools

CLAUDE.md / AGENTS.md per repo

Each client repo gets its own rules file with that client's decisions, so memory stays per repo by default.

Setup
1–2 hours per repo
Ongoing
update the files by hand after decisions, ~10 min a week per client

ArcRoom

Every agent reads the team playbook before it works and writes to it when it finishes. Nothing to update by hand.

Setup
2 min per repo (npx arcroom init --project <id>)
Ongoing
None

Per-repo files isolate just as well, but they cost tokens every prompt, each agent reads its own file, and nobody fills them unless someone writes them.

Where it stops

Shared knowledge across clients (your agency's own standards) has to be repeated in each project.

Try it in your project

Works with the agents your team already uses, and any agent over MCP or the HTTP API.

More in agencies and freelancers