We're not building agent memory

We're building the substrate for humans and AI to work together.

The phrase "agent memory" has always sat wrong with me. It frames the goal as helping the AI remember, as if the person is just a data source feeding the machine. On the contrary, I want to build the connective substrate between humans and artificial intelligence, where each can see, trust, and build on what the other did.

That's a different thing from a memory store, and it needs different properties.

It has to be legible: you can look at what the system knows, instead of hoping the right thing got embedded into a vector somewhere. It has to carry provenance: every fact tied to where it came from, so a person can trust it or push back. It has to let the two of you disagree: when the database says one thing and last week's meeting said another, the system shouldn't quietly pick a winner; it should surface the conflict and let a human settle it. And what comes out of all that has to be yours: your team's judgment, captured as structure that survives when you swap the model underneath, not baked into someone else's fine-tune.

I keep coming back to a conviction I've held for a while: getting AI into how people actually work is less a technology problem than a trust problem. People don't resist tools that make them more capable. They resist tools that feel like they're being replaced or overruled. A substrate where the human stays in authority (sees the sources, corrects what's wrong, and keeps the expertise) is the version of this that organizations will actually adopt. So it's the version I'm building.

To make it concrete, I built the foundation and put it in front of real data. You connect a Postgres database, read-only, and it reads your schema, drafts a typed model of your world, and syncs your rows in as objects. Then it fuses them with what your documents and conversations say about the same things. Ask about a customer and you get the database row and the line from last week's call where they hinted they might churn, with the source attached.

Here is a quick two-minute demo: https://www.loom.com/share/b5ee0abf39e54605ac89a66774fcf974

You can learn more about it here: https://mindgraph.cloud/

I think this matters because the rest of the field is quietly admitting the hard parts aren't solved: where a fact came from, what to do when two facts conflict, how to fuse a live operational database with everything unstructured around it.

Here's the honest part. This is early, alpha. And I'm not trying to win a signup race against the well-funded memory companies; I'd lose, and it's the wrong game anyway. What I want is to work closely with a small number of teams building agents on top of real operational data, in places where trust matters: legal, finance, anywhere a wrong or unsourced answer has consequences. I want to understand how you and your users actually work, and design this around that, with you.

So this is an invitation more than an announcement. If you're building something in that shape, I'd like to talk, and I'll go deep with the few it fits best. I'm happy to embed with a partner, to co-design with a group, to do the unglamorous work of making it real for one team before ten.

If that's you, reach out: DM me on LinkedIn: https://www.linkedin.com/in/shanrizvi/, or email shan@mindgraph.cloud