It is also the most dangerous.

Because the memory that makes your assistant feel like it finally knows you doesn't live with you. It lives on a server you'll never see, owned by a company whose incentives are not your incentives, governed by a privacy policy you didn't read and can't enforce. The single most intimate dataset you will ever generate — your questions, your fears, your half-formed plans, the things you'd never say out loud — is being assembled, indexed, and stored somewhere you don't control.

I think that's backwards. So my co-founders and I built the alternative.

Deja Vu is a local-first AI memory tool that gives your assistants persistent, private context — your data stays on your machine, never in someone else's cloud.

How I got here

I spent the early part of my career at Fidelity and Salesforce, and those two places taught me opposite halves of the same lesson.

At Fidelity, the entire institution is built on a single sacred principle: the asset belongs to the customer. The firm is a custodian. It holds your money, it does not own your money, and every system, every control, every line of compliance exists to keep that distinction absolute. You don't get to be casual about whose thing it is.

At Salesforce, I learned the inverse — the uncomfortable version. Whoever holds the data holds the relationship. The system of record isn't just storage; it's leverage. The company that owns your customer history owns a piece of your business, and they know it. That's not a flaw in the model. That is the model.

Now watch what's happening with AI memory and tell me it doesn't rhyme. Your memory — your preferences, your projects, your patterns, the accumulated context of your entire relationship with these tools — is quietly becoming the new system of record for your life. And by default, you are not the custodian. You're the product whose record someone else keeps.

I wanted to build the Fidelity version of AI memory in a world that's defaulting to the Salesforce one.

The bill is already coming due

If the risk of cloud-held memory felt abstract a year ago, 2026 has made it concrete with almost theatrical timing.

Earlier this year, one of the fastest-growing open-source AI agents in the world — tens of thousands of GitHub stars, adored by exactly the people who should know better — became the center of what security researchers are openly calling the first major AI agent security crisis. A single vulnerability allowed remote code execution from one malicious link. Tens of thousands of instances sat exposed on the open internet, leaking API keys, access tokens, and plaintext credentials. And it wasn't an outlier: researchers have catalogued a steady drumbeat of incidents over the past year that exposed the personal data of tens of millions of people across AI products. The lesson is structural, not incidental. When you funnel everyone's most sensitive context through centralized infrastructure, you don't just create a service. You create a target — a concentrated, irresistible one.

The regulators have noticed. Europe's data authorities issued a fresh reminder just days ago that persistent AI memory is profiling under existing law — which drags it into the territory of consent requirements and the right to be forgotten. New transparency obligations for these systems are weeks away from taking effect. There are active lawsuits alleging that tracking code was quietly embedded where private queries were being typed. And academic researchers have given the underlying problem a name: the personalization-convenience paradox. The feature users love most — the assistant that finally remembers — is precisely the feature they cannot audit, cannot constrain, and cannot fully trust.

That phrase has lived in my head for weeks, because it describes a trap, not a tradeoff. It says: to get the good thing, you must accept the unaccountable thing. You can have memory, or you can have control, but not both.

I refuse to accept that the two are mutually exclusive. The paradox isn't a law of nature. It's an artifact of where the data is stored. Move the storage, and the paradox dissolves.

What we actually built

Deja Vu is a memory layer that lives on your device, not on ours.

The architecture is deliberately, almost stubbornly simple. Your context is stored locally in SQLite — a plain, durable, inspectable database file sitting on your own machine. There is no account that has to exist for your memory to exist. There is no shadow profile of you accumulating on infrastructure you've never seen. The record is a file. The file is yours. You can open it, read it, back it up, move it, or delete it. Custody, in the Fidelity sense, sits where it always should have: with you.

And because it speaks MCP — the Model Context Protocol that has quietly become the connective tissue of the agentic ecosystem — that memory isn't trapped in one app. Your context travels with you, across the assistants and agents you use, instead of being held hostage inside whichever vendor you happened to start with. No vendor gets to own your relationship by owning your record. You bring your memory to the tools; the tools don't get to keep it.

That's the whole philosophy in one line: persistent context, private custody. The continuity that makes AI feel like it knows you, without surrendering the dataset that makes it know you.

Why now

This isn't a privacy lecture dressed up as a product. The market has already turned. Local-first stopped being a niche preference for the paranoid and became a mainstream position — analysts now treat on-device data handling as a primary axis of comparison between assistants, not a footnote. The personal-AI category is racing toward twenty billion dollars by the end of the decade. Memory is the feature everyone agrees is the differentiator. The only open question — the one the entire industry is now circling — is the one we built our company around: how deep does it remember, and can you trust it?

We think the honest answer to "can you trust it?" is: you shouldn't have to. Trust is what you fall back on when you've given up control. We'd rather just give you the control.

The last great shift in personal computing put a supercomputer in everyone's pocket. The next one is about to decide who owns the memory those computers keep about us. I don't think that should be a foregone conclusion decided by whoever moved fastest to centralize it.

Your AI should remember you. Nobody else has to.

Deja Vu is live. Your memory is waiting — right where it belongs, on your machine.

— building agentic memory at @useDejaVu