Open Brain: Building Your Own AI Memory That Isn't Locked to a Vendor

Every AI you use forgets you the moment you close the tab. That's not a technical limitation — it's a business model. Nate argues that per-vendor memory is a loyalty program dressed up as a feature, and walks through a DIY alternative: a Postgres + vector database you own, connected to every AI you use via MCP.

The Problem: Memory as a Moat

Every session starts the same way. "I'm working on a project where..." for the 15th time this month. You re-explain your job, your situation, what you already tried. It happens so often you don't notice anymore.

Some tools have added memory. Claude remembers some. ChatGPT remembers some. Helpful — but silo'd on purpose:

Claude's memory doesn't know what you told ChatGPT
ChatGPT's memory doesn't follow you into Gemini
Your phone app doesn't share context with your laptop

Every company is building a little memory silo, and they don't talk to each other because the more you tell one tool, the harder it is to leave.

The Analogy That Lands

Your dentist doesn't own your x-rays
Your bank doesn't own your transaction history
But right now, the AI that knows your job, your goals, your writing style — that company owns all of it

The moment you want to try a better model, you start from zero. That's a real switching cost, quietly working against you.

The Fix: Open Brain

Nate's solution: build your own memory layer and plug it into every AI via MCP.

Postgres database with vector embeddings — searches by what you meant, not just what you typed
Connected to every AI you use through the open MCP protocol
You own it — if a new model drops next Tuesday, you plug it in and get full context on day one

Practical numbers from the video:

Spec	Value
Setup time	~45 minutes
Monthly run cost	~$0.10 to $0.30
Storage	Postgres (you host it)
Protocol	MCP (open, multi-vendor)

At thirty cents a month you've liberated your own cognitive history from whichever AI vendor you happen to be using this week.

The Meta Point

The closing line is the real one. Nate says: you can take this video, feed it to an AI, and it will walk you through building this system. You don't even need his Substack quick-start. Just hand the description to a capable AI and it can scaffold the Postgres-plus-MCP setup for you.

That's the new baseline. Explainer-style content is now executable — consumed as instructions, not just information. The ceiling of what any individual can build has dropped dramatically, and vendor lock-in is a real-time choice you're making rather than a technical constraint.

Why This Argument Has Teeth

Memory is becoming the differentiator between "AI assistant" and "actual partner." A vendor's entire retention strategy will soon be built on how much they know about you. The more you invest in teaching one of them, the less rational it becomes to switch — even if a better model launches.

An open, portable memory layer flips that. Your memory is yours. The model is swappable. Switching cost goes to near zero, and the pressure on vendors shifts from lock-in to actual quality.

Key Takeaways

AI memory in closed products (ChatGPT, Claude, Gemini) is a business model, not a technical limit — it's a loyalty program dressed up as a feature
Build your own Postgres + vector memory and connect it to every AI via MCP to escape lock-in
Setup is ~45 minutes and runs for pennies a month — this is a DIY weekend project, not an infrastructure undertaking
Hand the video itself to an AI and it will walk you through building the system — explainer content is now executable
Your memory should not be some other company's moat

Resources

Model Context Protocol (MCP) — The open protocol for connecting memory across AI tools
pgvector — Postgres extension for vector similarity search
Nate B. Jones on Substack — Where Nate publishes the full quick-start guide