How it works
Three steps. Then it remembers.
Memexa is deliberately unspectacular to use: there's no new editor to learn and no method to follow. You keep working the way you do — just with memory.
Step 01
Connect
Create an account and connect your AI client — ChatGPT, Claude or Cursor. Sign in once in the browser, done. No installation, no technical skills.
Honestly measured: about five minutes. Then never again.
Step 02
Toss it in
Type notes, upload PDFs, save links, forward emails. Memexa reads everything — scans included, via OCR — and links it to projects, people and topics automatically.
No filing, no folder upkeep. The graph grows on its own.
Step 03
Ask
From now on your AI answers from your knowledge — in every connected client, across months and projects. Every answer names its source: which document, which page.
You can verify every claim. No guessing, no hallucination lottery.
For the technically curious
One address, every client
Under the hood, Memexa speaks the open Model Context Protocol (MCP). Your personal endpoint — paste it into any client, sign in once in the browser:
claude mcp add --transport http --scope user memexa https://brain.memexa.ai/mcp
Free — works on any Claude plan.
And then?
Connected? You don't need a manual — your first sentence to the AI is enough:
Remember this: I'm currently working on …
From there your brain grows in conversation — saving, asking and linking all happen through the AI itself.
Architecture details, the document pipeline and an honest comparison with other memory tools: technical page →
The quality
How good the answers get is partly your call
Memexa is the memory, not the head. We make sure the right passages are found — from notes, PDFs, scans, forwarded mail — and hand them to your AI with the source attached. What becomes of them is the model you're talking to. A strong model draws better conclusions from the same passages; a weak one just repeats them. That makes your choice of AI client part of the answer quality, and it's why we tie you to none of them: switch models when a better one arrives — your memory stays where it is.
Two things stay on our side regardless. First the groundwork: text recognition, distillation, linking and hybrid search all run at Memexa and are identical for every client — a weaker model doesn't get worse passages, it just gets less out of them. Second the ordering: when several hits sit close together, a cross-encoder sorts them. But what counts in the end is what your model does with the material.
The safeguards
What happens on the way there
Everything you upload sits on servers in Paris. Every leg of the way is encrypted — including the one between application and database, and there with certificate and hostname verified, not just "encrypted somehow". The database itself has no public address; it sits on a private network you can't reach from outside.
The separation between accounts isn't in the application code, it's in the database: with no user context set, PostgreSQL hands out no rows. A mistake in a single query therefore can't expose someone else's data — the database returns nothing at all.
Your AI clients don't get a master key, they get an access you approve beforehand and can disconnect any time. The browser clipper and full access are separate: the clipper may file things, not search. And when you leave, files are deleted first and records second — if the first step fails, the whole run aborts rather than leaving remnants behind.
If you want the long version — every processor with name, seat and place of processing, plus the list of what we don't claim: trust page →
The best time: before you forget it again.
Set up in 5 minutes, no credit card.