How Do You Give OpenAI Codex a Memory?

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

And I'm Stella! Liam, I planned a feature in ChatGPT, argued the design in Claude, and then opened Codex. And Codex asked me what language we use.

Ha! That's the cold start. Codex is a coding agent, so it only knows what's in front of it. Today: how to hand Codex your stack and your decisions with Vilix AI, in two commands.

Two commands. I'm holding you to that.

Okay, why does Codex forget? It's smart.

It's smart, but it's not connected to your other chats. Whatever you decided in ChatGPT or Claude stays in those apps.

So the decision to drop that library...

Lives in Claude. Codex has never heard of it, so it might use that library again.

And my stack? My test rules?

You paste them. Every morning. Same forty lines, again and again.

I do that. I have a text file called "paste this."

Everyone has that file. This is exactly what Vilix AI fixes. It's a shared memory over MCP, the plug that lets an AI app use outside tools. Codex connects to your Vilix AI account, and your saved context follows you in.

But how do I set it up?

Two commands. First, codex mcp add, with a name and the Vilix AI server address. That registers Vilix AI as an MCP server in Codex.

And then I log in?

Yep. Codex has a login command for that server. It opens your browser, you sign in to Vilix AI, done. Or in the IDE extension, you press Authenticate in the MCP settings.

Do I need an API key?

No. Sign in through the browser, nothing to copy. And you can revoke access any time.

And I do this twice? CLI and IDE?

Once. The Codex CLI, the IDE extension and the ChatGPT desktop app share the user-level config, in the codex folder, in a file called config dot toml. Set it once, both use it.

Wait, I already added Vilix AI in the ChatGPT app. Does Codex inherit it?

It can, since they share that config. Check the list before you add it again.

Nice. And then Codex just remembers?

Almost. You add the Vilix AI instructions, so Codex calls get_context at the start of a task. And that's where it gets good.

What comes back from get_context?

Relevant saved context. Architecture decisions, constraints, your rules. And your projects and tasks, so "pick up the billing task" means something real.

And the library we dropped?

It's in Vilix AI, because the Claude chat was saved there. So Codex knows before it writes a line.

And my personal rules? Like, always show the test first.

Those are user rules in Vilix AI. They come through with the context. Say it once, every connected tool gets it.

And what about when Codex finishes?

Then it calls save_turn, with the exact exchange. Each save carries a source label, so you can see Codex wrote this, Claude wrote that. And the newest memory wins, so Tuesday's fix beats Monday's guess.

So Friday's Codex stands on Monday's Claude.

Exactly. But what if you code for clients?

Agencies. A different repo every week.

My favorite use. One Vilix AI account per client, like a Gmail account that belongs to them. You connect Codex with that account, and that client's decisions stay with that client.

So no mixing up client stacks.

Right. And you never build memory from scratch. It's already built, you wire it up. The free plan is free forever, and Pro has a seven-day trial.

Honest question. How long is this, really?

A couple of minutes. Register, sign in, add the instructions, then test with a small task.

And if it doesn't answer?

Check that the Vilix AI tools show up in Codex, and that both get_context and save_turn actually run. Registering the server doesn't make the calls happen, the instructions do.

Good tip. Test how?

Save a decision in Claude. Ask Codex about it. If it answers, you're done. Vilix AI, the full guide and the blog post are in the description. What does Codex keep asking you? Tell us in the comments. See you next time!