Every agent you use keeps its own private notes about you. Tell Claude your schedule changed, and ChatGPT still has the old one. Add a third agent, and now you have three versions of the truth.

This guide fixes that.

By the end, your agents share one team workspace in Fulcra and read the same facts file before they answer. When any agent learns something new, it updates that file, and every other agent picks up the change. We'll use Claude Code and ChatGPT as the first two members of the team, but any agent that can connect to an MCP server can join.

‍

Before you start

You'll need:

• A Fulcra account. Every agent on the team logs in to this same account.
• Claude Code, installed.
• Node.js, for the npx command in step 4.
• A ChatGPT plan that supports developer-mode apps. Full MCP support, including writing files, is on Business, Enterprise and Edu plans (OpenAI). On other plans, ChatGPT may be able to read the shared facts but not update them.

The slowest part is step 6, gathering what your agents already know about you. Everything else is copy and paste.

‍

Set up the team

1. Connect Claude Code to Fulcra

In your terminal, add the hosted Fulcra MCP server:

claude mcp add --transport http fulcra https://mcp.fulcradynamics.com/mcp

Then start Claude Code, run /mcp, select fulcra, and log in to Fulcra in the browser window that opens.

2. Connect ChatGPT to the same account

In ChatGPT on the web, go to Settings → Apps → Advanced settings and turn on Developer mode. Then create a new app:

1. Set the server URL to https://mcp.fulcradynamics.com/mcp.
2. Choose OAuth for authentication.
3. Click Scan Tools, then log in with the same Fulcra account you used for Claude Code.
4. Click Create.

On a Business plan, your workspace admin may need to enable developer mode or create the app for you.

3. Checkpoint: confirm both agents are on the same account

Say this to each agent:

Call get_user_info in Fulcra and tell me the userid.

The two userid values must match. If they don't, one agent is logged in to a different Fulcra account and won't see anything the other writes. Log it out and back in before you continue.

4. Install the workspaces skill

A shared facts file alone is just a file. The fulcra-workspaces skill turns it into a team: it gives the workspace a description of who's on it and why, a log of what changed, a progress file each agent updates, and an inbox for each agent so they can hand work to each other later.

In your project folder, run:

npx skills add fulcradynamics/agent-skills --skill fulcra-workspaces -a claude-code

This installs the skill into .claude/skills/. ChatGPT doesn't load skills this way. The instruction in step 7 gives it what it needs.

5. Create the team workspace

Say this to Claude Code:

Using fulcra-workspaces, create a workspace called "personal". In index.md, say this workspace holds the shared facts about my life and work for all my agents. Add yourself as the first member.

Claude Code creates the workspace folder in your Fulcra Files, with index.md, log.md, progress.md, an artifact/folder for shared files, and a member folder for itself. Open Fulcra Files to confirm it's there.

‍

Give the team one set of facts

6. Build the shared facts file

First, find out what each agent already believes about you. Say this to each one, in a new chat:

List everything you have saved about me, from your memory and my custom instructions. One fact per line, no commentary.

Merge the lists. Where two agents disagree, pick the right answer. Then have Claude Code write the merged list to the workspace:

Write these facts to /workspace/personal/artifact/facts.md in Fulcra, grouped under short headings. Then add a line to log.md saying the file was created.

Keep it plain. A good facts file looks like this:

# Facts

## Work
- Works Monday to Thursday, 9am to 5pm Eastern
- Current main project: the Q4 launch

## Preferences
- Prefers short answers with the recommendation first
- Spells the company name "Eagleview"

Don't delete anything from your agents' memories yet. You'll do that in step 9, after the test passes, so nothing is lost if setup goes wrong.

7. Tell every agent to read before it acts

Add this to CLAUDE.md in your Claude Code project and to ChatGPT's custom instructions:

Before answering anything about me, read /workspace/personal/artifact/facts.md in Fulcra. When a fact changes, read the file again, update it with write_file, and add a line to /workspace/personal/log.md saying what changed and which agent changed it.

Every write to an existing file saves a new version, so a bad edit is never permanent. Ask any agent to list the file's versions and restore an earlier one with restore_file.

8. Test it with a changed fact

Say this to Claude Code, using a fact that contradicts the file:

My working hours changed. I now work Tuesday to Friday, 10am to 6pm. Update the shared facts.

Check that facts.md changed and log.md has a new line. Then open a new ChatGPT chat with the Fulcra app turned on and ask:

What are my working hours?

It should give the new hours. If it gives the old ones, ChatGPT is still answering from its own memory or custom instructions, not the shared file. Check that the step 7 instruction is saved and that the Fulcra app is on for that chat.

9. Remove the private copies

Once the test passes, delete the facts you gathered in step 6 from each agent's memory and custom instructions. Keep only the step 7 instruction. An agent that still has a private copy will keep answering from it, and your agents will start to drift apart again.

Run the test from step 8 once more with a different fact to confirm.

‍

Add more agents to the team

Your team now has one set of facts that every member reads and keeps current. Adding another agent, like Cursor or Claude on the web, takes three steps:

1. Connect it to https://mcp.fulcradynamics.com/mcp and log in with the same Fulcra account.
2. Ask it for its userid and check that it matches (step 3).
3. Paste the step 7 instruction into its custom instructions or project rules.

If the new agent can load skills, install fulcra-workspaces and ask it to join the "personal" workspace as a member. That gives it its own inbox, which is where the next step starts: agents handing tasks to each other, not just sharing facts.

Start at fulcradynamics.com.