What gets shared and what does not
An agency writes for ten clients in ten voices. One client never wants the word “cheap”, another has a product name with fixed capitalization, a third writes in a formal register. As of November 2026, a shared Claude project splits like this:
| Part | Shared with the team? | Use it for |
|---|---|---|
| Project instructions | Yes, with project members | The hard rules of the client’s voice |
| Project knowledge | Yes | Brand guide, glossary, approved examples, briefs |
| Project memory | Not stated; Claude’s memory is per user | What came up in chats in that project |
| Account memory | No, per employee | Personal working style, not client rules |
| Your chats | No, each person’s own | Drafts and iterations |
Sharing requires Team or Enterprise. Claude Team covers 2 to 150 people at $20 per standard seat per month billed annually or $25 monthly, excluding tax, and it lists memory and project sharing as included. Project memory and chat search also carry some context from one chat to the next, but that is per user and not something the agency can manage. What every colleague needs in every chat belongs in the instructions or project knowledge.
A setup per client
- One project per client, on Team or Enterprise, because only those plans let you share a project with colleagues.
- Instructions: the hard rules. Form of address, banned words, name spellings, length, what may never be promised.
- Knowledge: the material. The brand guide, three to five texts the client approved, the current brief.
- Turn off Use account memory under Memory on the project page. The project then uses only its own memory, not what an employee did for another client in loose chats.
- Share per client team, with Can view (see and chat) or Can edit (change instructions and knowledge). Give Can edit to whoever owns the client relationship.
- Every client correction goes into the project. If the client says “never this word again”, put it in the instructions. An employee’s memory may remember it too, but then it is not an agreement the whole team shares and can check.
Here is what the instructions for one client can look like (a made-up example):
Client: Example Bakery
Address the reader as "you", casual.
Never: "cheap", "deal", exclamation marks.
Product names: Sourdough One, always capitalized.
Length: posts up to 80 words.
Examples: see "approved-2026.md" in project knowledge.
Check: never mention another client or use words from another client's list.
When the client corrects a draft, update this file the same day. When a freelancer joins as a user in your organization, add them to the project with Can view; the rules travel with the project, not with the person who wrote them.
For the step-by-step sharing flow, see how to share Claude Projects with your team.
The misconception: “the AI will get to know our clients”
Claude’s memory is per user. What the AI picks up about a client from your conversations does not help your colleague, and owners can’t view or edit a user’s individual memories. A project also pulls in your account memory from chats outside projects by default. If you wrote a quick post for client A outside a project this morning, that voice can bleed into client B’s project.
When people rotate on an account, or a freelancer helps out, knowledge that lived only in someone’s memory leaves the table. And a freelancer can only join a shared project as a user in your organization.
What this means for an agency
A client’s agreements are agency knowledge, not one employee’s. One project per client handles that inside Claude. But agencies rarely work in one tool: the copywriter in Claude, the social team in ChatGPT, the account manager in Copilot with the client’s emails. Each tool keeps context in its own place, and a correction handled in one is unknown to the next. Sharing a Copilot notebook shares the linked files, provided you have sharing rights on each, but Microsoft does not say whether chats or memory travel with it. For the broader picture, see how to share AI memory across a team and does Claude have memory.
Where Bombos fits
Bombos is a plugin that gives every AI tool in the agency, Claude, ChatGPT, Copilot and your own agents, one shared memory. A client’s voice and agreements become rules that apply only within that client, and a correction recorded in one tool is known to the next. Each answer shows what it rests on. It runs in the cloud or on your own server, and one person can try it first for their own agents.
Related questions
Can a freelancer join our shared Claude project?
Only as a user in your organization. On Team and Enterprise, projects are shared with other users in your organization, as Can view or Can edit.
Why not let Claude's memory learn how each client writes?
Because that memory belongs to one employee. A colleague does not get it, and owners can't view or edit a user's individual memories. What the whole team needs belongs in project knowledge.
How many projects can we create?
Free accounts are limited to five; paid plans allow more. To share projects with colleagues, an agency needs Team or Enterprise.
Can an admin stop people from sharing client projects?
Yes. Admins can turn off project sharing, and on Team and Enterprise the owner decides whether memory is available at all.
