What your workspace knows does not have to stay in one window. Point another
assistant at it and that assistant can search what you have gathered, answer
from it, and add to it.

The memory stays where it is. Nothing is copied into the other app, nothing is
installed on your machine, and everything the assistant does is on a record you
can read afterwards.

## Four steps

**1. Have a workspace.** If you do not yet, [set one up](/docs/workspace/setup).
It takes about five minutes, most of it waiting. Give it something to read
before you connect anything, or the assistant will find an empty memory and say
so.

**2. Add the memory to a client.** A client is the access you hand out.
Open **Applications**, find your workspace under Memory, and install it.
The [Applications page](/docs/tools/applications) covers this for every kind of
application, memory included.

**3. Choose Read or Full.** Read lets the assistant search and recall. Full also
lets it remember new things, link notes, and forget. You can change it later,
and the change takes effect on the assistant's next call. See
[Permissions](/docs/tools/permissions).

**4. Connect the assistant.** One address, the same for every app:

```
https://gbrain.io/mcp
```

[Connect an assistant](/docs/tools/assistants) has the exact steps per app,
including where each one hides its MCP setting.

## Per assistant, the short version

Full steps are on the [Connect an assistant](/docs/tools/assistants) page. The
one thing that differs per app:

- **Claude Code.** `claude mcp add gbrain -t http https://gbrain.io/mcp`, then
  sign in when it asks. Your memory shows up as tools named after the server you
  gave it.
- **Claude, web or desktop.** Add the address as a connector under Settings and
  approve the sign-in.
- **ChatGPT.** Add the address as a connector and approve the sign-in. Deep
  research works against your memory too, because it looks for the same search
  and fetch tools your workspace already publishes.
- **OpenClaw.** Two settings matter. Its default transport is the older SSE, so
  set `transport: "streamable-http"`, and use `auth: "oauth"`. If several people
  talk to one OpenClaw, add `oauth.identity: "per-requester"` so each person
  signs in as themselves and the record shows who asked.
- **Hermes.** Add the address under `mcp_servers`. It signs itself in.

## Try it

Once it is connected, these are the asks that prove it:

- *"Search your memory for what we decided about pricing."* Reading.
- *"Remember that Ana runs procurement."* Writing, so it needs Full.
- *"What do you know about this company already?"* The reason to connect at all:
  the answer comes from your material, not from the model's guesswork.

An assistant will not reach for memory on its own for every question. Saying
"search your memory" is the difference between an answer from your notes and an
answer from general knowledge, and the two sound equally confident.

## Good to know

**One workspace per client.** A client holds one workspace's memory, so a second
workspace means a second client. Tools keep their plain names that way, and an
assistant never has to guess which memory it is writing to.

**Off until you connect it.** Nothing listens on your workspace until you add
its memory to a client. Take it off every client and it switches off again.

**Everything is on the record.** Every search, every note written, which
connection did it, and what it cost, on the [Activity page](/docs/tools/activity).

**The assistant cannot widen its own access.** If it needs Full and has Read, it
will say so and hand you the page to change it. Only you can change it.