MemoryBase

Every time you use AI to write, code, plan, or think something through, you're building context. Your habits. Your patterns. Your tone.

Across a normal week:

• ChatGPT learns about your personal admin, how you structure emails to your team, and your writing tone of voice.

• Claude learns how you structure code and how you build projects, and how to use your collection of skills.

• Every new AI app captures another slice of you that nothing else can see.

None of them share. Your context splits across these tools and never adds up to a single version of you. That's the surface problem you can feel every day. The deeper one is underneath: none of that context actually belongs to you.

Memory is becoming the next lock-in

Over the last two years, every major AI tool shipped memory.

Three years from now, switching between AI tools will cost you everything you've taught the AI. Not because the AI tools are bad. Because the memory is bolted to the platform, not to you.

We didn't accept this for documents. We didn't accept it for email. We shouldn't accept it for the substrate that makes AI useful.

Andrej Karpathy called the real work of getting good output from an LLM context engineering, not prompt engineering: "the delicate art and science of filling the context window with just the right information for the next step" (Karpathy on X, June 2025). If context is the lever for AI performance, the lever has to sit in the hands of the person doing the work, not the platform serving them.

What AI memory should be

A few principles I think are non-negotiable:

1. It follows the user, not the app. Your context is a property of you, not of whichever tool you used last.

2. You own it. Browse, edit, export, or delete any of it. No part is vendor-locked.

3. It's one continuous thread. Across every AI you connect to, not five disconnected stores that contradict each other.

4. You decide what gets shared, with which AI, for what task. Context belongs to you. The provider has to be invited.

That's the layer that should exist. Memory belongs to the user. Everything else is a feature inside someone else's app.

Why this has to come from outside the labs

The model providers aren't going to build this themselves. They have direct incentives to keep your context inside their tools, it's how a paid AI tier earns its keep, and how a user becomes too sticky to leave for the next model.

The layer that follows you, that you control, that any model can read from has to come from outside the model labs. That's the bet I'm making with MemoryBase.

Plenty about the execution isn't settled (auto-organizing thousands of conversations is closer to a search-ranking problem than a database lookup, and we know which categories we still fumble on), but the principle is still non-negotiable.