User-Owned AI Context: The Real Privacy Story for the AI Era

Question: what do your AI tools currently remember about you?

If you use Claude, ChatGPT, Cursor, Codex, Perplexity, and a couple of custom GPTs, the honest answer is: you don't know.

Each vendor maintains its own memory store. Each has different retention policies. Each has different export options. Each has different access controls. You can ask ChatGPT what it remembers about you and get one answer; you can ask Claude the same question and get a different one. There is no consolidated view.

This is not a feature. This is a structural problem.

What "AI privacy" usually means

When AI companies talk about privacy, they usually mean two things:

1. "We don't train on your data." (Mostly true, often with caveats.)
2. "We have controls so you can delete things." (Usually true, often hard to find.)

These are necessary. They are not sufficient. They both assume the vendor remains the steward of your data. The privacy choice you're offered is "trust us with your context, with these guardrails."

The deeper privacy question isn't about training or deletion. It's about ownership. Who holds the canonical record of your AI context? Today, the answer is "each vendor holds a fragment." Tomorrow's better answer is "you do."

What user-owned context actually means

A user-owned context layer has four properties:

1. You can see all of it. Not just a vendor-specific subset. The complete graph of what's been captured, where it came from, and what's been derived from it.

2. You can revoke any of it. Not by submitting a deletion request to a vendor and waiting; by removing the memory from your graph, immediately.

3. You can export all of it. Not just an opaque archive; a structured dump you can move to a competing service if you want.

4. You control what each AI tool can see. Scope, per destination, per task. Personal context stays out of work. Work stays out of personal. Sensitive projects stay out of low-trust tools.

The first three are about data sovereignty. The fourth is about routing. Together, they constitute what we mean by "user-owned context infrastructure."

How Memplex is built around this

Memplex's design choices all flow from the ownership principle.

• Local-first capture. The Desktop Agent runs on your machine. Filesystem watchers ingest from your sources locally. Your raw context never leaves your machine unless you choose to sync it.

• Source ACL inheritance. The Memplex graph mirrors the permissions of the sources. If you lose access to a Slack channel, the derived memories become inaccessible automatically.

• Per-destination scoping. Every AI tool gets exactly what its policy and your scope rules allow, nothing more.

• Full lineage. Every memory has a source link, a confidence score, a lifecycle state. You can verify everything.

• Export and revocation. The graph is yours; you can dump it, prune it, or delete it at any time.

None of this is a premium feature. It's the architecture.

Why vendors won't build this

Every vendor has incentive to keep your context inside their walls. Memory is the moat. Letting you take your context with you when you switch to a competitor is, from their perspective, value leakage.

Vendor-owned memory is rational for the vendor. It is bad for the user.

The privacy story for the AI era can't be "trust each vendor to do the right thing with your data." There are too many vendors, the data flows are too complex, and the incentives are too misaligned. The story has to be infrastructure: a layer that you control, that the vendors plug into, where the data sovereignty defaults are inverted.

That's what we're building.