The portability test for your company’s AI memory

If your AI vendor disappeared tomorrow, what would you actually still own? That may be the simplest test for whether you really own your company’s AI memory.

In my last post, I asked whether companies should build or buy the emerging AI agent layer. The discussion pushed the question one level deeper: what does “owning the memory” actually mean? It definitely does not mean exporting a folder full of chat transcripts.

A useful agent gradually accumulates something much more valuable: what your company knows, who knows what, how decisions get made, which sources can be trusted, what an employee prefers, how workflows actually operate, what exceptions have been made before (this is a big one), and what the agent is allowed to see and do.

That context becomes more valuable the longer the system runs. So I think companies need a portability test. If you decided tomorrow to replace OpenAI with Anthropic, Google, xAI, an open-weight model, Instinct, Muse, Grok Bot, or something that doesn’t exist yet, could you move memory, permissions, provenance, workflows, and institutional context?

If the answer is no, you may not own the intelligence layer. You may be renting it. And I don’t think the answer is therefore “build everything yourself.” Models will change. Vendors will change. Infrastructure will change. New agent architectures will appear.

The goal should be to own the assets that become more valuable with time, while renting the layers that are becoming commodities. That probably means rent the model, buy tools where buying is faster, own the memory / context, govern the authority, and keep the architecture portable.

And because this market is changing so quickly, this cannot be a decision you make once. I think companies will need something like a quarterly AI architecture review:

What do we still need to own?

What can we now buy?

What has become a commodity?

Where are we becoming locked in?

What has appeared in the last 90 days that changes the answer?

We barely talked about personal AI agents a year ago. Whatever the right architecture is today, I would assume it will be wrong eventually. The real advantage may not be choosing the right stack. It may be building a company that can keep changing it.