Three steps. Then your AI 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.

  1. 01 Connect
  2. 02 Save
  3. 03 Ask

Example scene: a scanned insurance PDF lives in Memexa. Asked about the deductible, the AI answers with the amount and gives the citation (document, page 4) while the passage is highlighted in the scan.

Scripted illustration of real tool calls.

  1. 01 Step

    Connect

    Create an account and connect your AI app: Claude, ChatGPT or Mistral Vibe. Sign in once in the browser, done. No installation, no technical skills.

    Honestly measured: about five minutes. Then never again.

  2. 02 Step

    Save

    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 connections form on their own.

  3. 03 Step

    Ask

    From now on your AI answers from your knowledge, in every connected AI app, across months and projects. Every answer gives its citation: which document, which page.

    You can verify every claim. No guessing, no made-up answers.

Which AI apps work, and how

You paste one address into your AI app and sign in once in the browser. From then on the AI app reads and writes your memory.

The address for every AI app

https://brain.memexa.ai/mcp
  • Claude (claude.ai, desktop app)

    Plan needed
    Every plan. Free: exactly one custom connector; Pro and Max: several
    How
    Settings → Connectors → Add custom connector, then paste the address. Open the dialog directly
  • ChatGPT

    Plan needed
    Free plan included (tested by us on Free and Plus)
    How
    chatgpt.com/plugins → Add → Add custom MCP server → enter name, server URL and OAuth → Create as plugin
  • Mistral Vibe (formerly Le Chat)

    Plan needed
    Every plan, free included
    How
    Create a custom MCP connector and paste the address.

As of October 2026, partly tested in real accounts. Menu names in the AI apps can change.

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 memory grows in conversation: saving, asking and linking all happen through the AI itself.

For Claude Code and Cursor

Claude Code
Every Claude plan
Cursor and other MCP tools
Depends on the tool
claude mcp add --transport http --scope user memexa https://brain.memexa.ai/mcp

Run it in the terminal. You sign in once in the browser.

For the technically curious

Under the hood, Memexa speaks the open Model Context Protocol (MCP) over HTTP. Sign-in runs through OAuth 2.1 in the browser; there are no tokens to copy. For the background, read What is MCP?; the steps for Claude are in Set up Memexa in Claude.

Architecture details, the document pipeline and an honest comparison with other memory tools: technical page →

How good the answers get is partly your call

Memexa finds, your model thinks

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 their citation. What becomes of them is down to the model you're talking to: a strong one draws better conclusions from the same passages, a weak one just repeats them.

You're not tied to any model

That makes your choice of AI app part of the answer quality, and it's why we tie you to none. Switch models when a better one arrives. Your memory stays where it is.

The groundwork is the same for everyone

Text recognition, distillation, linking and search run at Memexa, the same for every AI app. A weaker model doesn't get worse passages, it just gets less out of them. When several hits sit close together, an additional model sorts them by relevance.

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: it only hands out records once it's clear which account is asking. A mistake in a single query therefore can't expose someone else's data: the database returns nothing at all.

Your AI apps 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, with every processor's 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

Try free for 14 days

Then from €2.99/month · cancel anytime before · card required