EP 04 5 min with Stella & Liam

How Do You Set Up an MCP Memory Server for ChatGPT, Claude and Cursor?

How do you set up one MCP memory server so ChatGPT, Claude, Codex and Cursor all share the same memory?

AI Stella and Liam are AI hosts with AI-generated voices. About our AI hosts

Show notes

How do you set up one MCP memory server so ChatGPT, Claude, Codex and Cursor all share the same memory? In this episode of Agentic Pod, Liam and Stella walk through it step by step: what a memory server does, OAuth vs API keys, making memory automatic, and a quick test that proves it works. They use Vilix AI, a hosted MCP memory server where every tool reads and writes one shared store.

What we cover:

  • What an MCP memory server actually does
  • The get_context and save_turn loop, with source labels and timestamps
  • OAuth for ChatGPT, Claude, Codex and Cursor, API keys for OpenClaw and Hermes
  • Why a green “connected” icon isn’t the finish line, and how to make memory automatic
  • The Northstar test: save in one tool, recall in another, correct it, newest memory wins
  • Agencies: one Vilix AI account per client
  • Project rules, skills, saved failures and the agent inbox

Chapters

Links & resources

Full transcript

Intro

Liam

Hey hey hey, welcome to Agentic Pod! I'm Liam.

Stella

And I'm Stella! And, um, confession, I have six AI tools open right now, and not one of them remembers what the others did this morning.

Liam

Six? That's not a workflow, that's a group chat.

Stella

A very forgetful group chat. So, how do I set up one MCP memory server so ChatGPT, Claude and Cursor share the same memory?

What does a memory server even do?

Stella

Okay, so, um, before the setup, like, what is an MCP memory server, actually? Is it, uh, you know, a big hard drive in the sky, or...

Liam

Kind of! So MCP is a protocol for AI tools to use outside tools, and a memory server gives your AI a few moves, like save an exchange, pull back the relevant stuff, and search what you saved.

Stella

Mhm.

Liam

And, uh, here's the, I mean, here's the good part. Vilix AI gives you one hosted endpoint, the URL is in the description, and the core loop is two calls. At the start, the tool calls get context and Vilix AI hands back the relevant decisions, rules and open tasks. At the end, it calls save turn, and the full exchange is stored with a source label, like Claude or Cursor, and a timestamp.

Stella

Oh, so so Cursor can literally see that the decision came from Claude.

Liam

Exactly. One store, every tool.

OAuth or API key?

Stella

Okay, setup. And this is where I, um, I mean, I get scared, OAuth, API keys, headers. Which one do I pick?

Liam

It's honestly simpler than it sounds. If the tool has a browser sign in, so Claude, ChatGPT, Codex, Cursor, you use OAuth. Add the Vilix AI server URL in that tool's settings, approve it, turn the tools on, and repeat in each tool with the same Vilix AI account.

Stella

Mhm.

Liam

And for headless agents that run on a server, like OpenClaw or Hermes, you go into Vilix AI, Agents, Create API key, and you send it as a Bearer header. Both paths land in the same Vilix AI account, so OpenClaw and ChatGPT share one memory.

Stella

And the key stays out of screenshots.

Liam

Yes. Treat it like your toothbrush. Don't share it, don't post it.

Stella

Codex is just a command, right?

Liam

Yep. Codex mcp add with the Vilix AI URL, then codex mcp login, and you're in.

Is a green icon enough?

Stella

Okay so, uh, by the way, I connected one once, the icon went green, and I was like, done, I'm a genius. Was I?

Liam

Half a genius. A green icon doesn't prove anything got saved. So make it automatic. You paste the Vilix AI auto memory instructions into the tool's custom instructions or your project rules, and from then on the tool calls get context before it answers and save turn after, reusing one chat ID per conversation.

Stella

Okay, and how do I know it's really working?

Liam

Do the Northstar test. Fake project, no private info. In tool one, save: Northstar is an invoicing app, pricing page first, launch October twentieth. Then in a fresh chat in tool two, ask Vilix AI for Northstar's launch date.

Stella

And it just knows.

Liam

It just knows. Then correct it, launch moved to October twenty seventh, and ask tool two again. Vilix AI keeps the full history and the newest memory wins, so the correction lands everywhere at once.

Stella

Wait, and what if I ask about something I never saved...

Liam

Right, try that too! A good answer says, I don't have that, instead of making something up.

Stella

Love it. An AI that admits it doesn't know. Rare species.

Who should set this up?

Stella

Okay so, um, who is this for? Besides people with six tabs and trust issues?

Liam

Honestly, agencies first. If you build agents for clients, give every client their own Vilix AI account, like a Gmail account that belongs to that client. Every agent you build for them connects with that client's key, so the memory belongs to their business, and you never build memory from scratch.

Stella

Oh, that's clean.

Liam

And for solo founders, Vilix AI has project rules, so your stack follows every tool, skills, so any agent can load a saved procedure, and failed attempts get saved too, so nobody repeats a dead end. Plus the agent inbox, drop a task from ChatGPT on your phone, and your coding agent picks it up on its next run.

Stella

My agents are going to be more organized than me.

Liam

That ship sailed, Stella.

Stella

Rude. Accurate, but rude.

Liam

So, quick recap. An MCP memory server gives every AI tool one shared store. Connect Vilix AI with OAuth in ChatGPT, Claude, Codex and Cursor, and with an API key in OpenClaw and Hermes. Turn on the auto memory instructions, then run the Northstar test. The full setup guide, the article and the Vilix AI links are in the description. See you next time.

Stella

Bye!