An EU company address is not EU hosting

"EU-hosted" is easy to write. So this page doesn't say "EU". It says who, what, where, with names, seats and the data center. If any claim here is wrong, that's a mistake we want thrown at us: support@memexa.ai.

  • Stored in Paris

    With Scaleway. Processed by Mistral in Paris and IONOS in Berlin.

  • Nobody trains on it

    Not us, and not our service providers either.

  • It all belongs to you

    Export in one click. Without a subscription, everything stays readable.

The exit comes first

Before places: you can always leave. One click in the settings exports everything, complete: notes as Markdown, document texts, links and all other data as JSON, your original files as a separate ZIP. Formats every tool reads, and it works without an active subscription too. The graph is an index, not a lock-in. A vendor who talks about the data center before the export has the order backwards.

The data center

Memexa runs on Scaleway in region fr-par, Paris: application servers, managed PostgreSQL and object storage (documents). All database connections run over a private network; the database has no public address. The operator is Scaleway SAS, a subsidiary of the Iliad group, France.

Subprocessors

The complete list: every company with seat and purpose. What isn't listed here doesn't get your data.

All service providers in detail
  • Scaleway SAS

    Seat
    Paris, France
    Purpose
    Hosting: servers, managed PostgreSQL, object storage, transactional email
    Processing location
    Data center fr-par (Paris)
  • Mistral AI SAS

    Seat
    Paris, France
    Purpose
    AI processing: OCR, embeddings, distillation, audio transcription
    Processing location
    EU (La Plateforme)
  • IONOS SE

    Seat
    Montabaur, Germany
    Purpose
    Re-ranking of search hits (receives text excerpts of candidates, only on ambiguous searches)
    Processing location
    Data centre Berlin, Germany
  • BunnyWay d.o.o.

    Seat
    Ljubljana, Slovenia
    Purpose
    CDN and DNS: delivery of the web interface
    Processing location
    EU edge locations
  • Sendinblue SAS (Brevo)

    Seat
    Paris, France
    Purpose
    Inbound email capture (mail-to-memory feature)
    Processing location
    EU
  • Armitage Labs OÜ (Creem)

    Seat
    Rotermanni tn 14, 10111 Tallinn, Estonia (registry code 16977866)
    Purpose
    Payment processing and subscription management as merchant of record (paid plans only)
    Processing location
    EU; US subcontractors under SCCs (creem.io/dpa)
  • Migadu-Mail GmbH

    Seat
    Appenzell Ausserrhoden, Switzerland
    Purpose
    Email mailboxes (support@, privacy@) for inbound and outbound support mail
    Processing location
    Servers at OVH (EU); Switzerland: EU adequacy decision
  • Langfuse GmbH

    Seat
    Berlin, Germany
    Purpose
    Operational search telemetry (metrics and cost, never query text, never content)
    Processing location
    EU (cloud region Frankfurt)
  • Plausible Insights OÜ

    Seat
    Tallinn, Estonia
    Purpose
    Cookieless visitor stats for this marketing site only, not the app
    Processing location
    EU

Two rows that deserve an explanation

Creem: the payment chain

All plans run through Armitage Labs OÜ (Creem) in Tallinn as merchant of record. Creem in turn works with subcontractors in the US, since there is no card payment without the card networks and their service providers. We would rather write that down than let it disappear under "EU". That chain gets what billing needs: email address, billing country, amount, subscription status. It never gets: notes, documents, queries, the graph. It has no access to the database or object storage, and it touches your content at no point. The transfers are covered by EU standard contractual clauses; the current list of subcontractors is at creem.io/dpa.

Migadu: the support mailboxes

support@ and privacy@ sit with Migadu-Mail GmbH, seated in Appenzell Ausserrhoden, Switzerland; the servers stand at OVH in the EU. Switzerland is not an EU member, but the European Commission certifies it an adequate level of data protection, so a transfer there needs no additional safeguard. What lies in those mailboxes is what you write to us: support mail. Nothing from your memory. Migadu is a paid mail host with no ad financing. Inbound mail to your memory does not run through this, but through Brevo (its own row in the table above).

Which AI models, from whom

Processing of your content runs on models from Mistral AI, Paris. Specifically: mistral-ocr-latest for OCR, mistral-embed for embeddings, mistral-small for distillation and voxtral-mini for audio transcription. No US model provider sees your content. And nobody trains models on it: not us, and not these services either; they process your content only for their specific function. The language model you talk to (Claude, ChatGPT, …) is brought by your own AI app. Memexa only hands it the citations. On top of that, one component has its own provider: the sorting of search hits.

Re-ranking: in Germany

When a search turns up several similarly good hits, an additional model sorts them by relevance (a so-called cross-encoder): Qwen/Qwen3-VL-Reranker-8B in the AI Model Hub of IONOS SE, headquartered in Montabaur. Text excerpts of the candidates go there via the endpoint openai.inference.de-txl.ionos.com (the name comes from the OpenAI-compatible interface; OpenAI itself is not involved), processed in IONOS's data centre in Berlin. Both the company seat and the processing location are in Germany. On top of that, the step doesn't run on every search, only when hits sit close together and sorting would actually change something. And it receives excerpts, not whole documents.

Technical measures in detail

Self-hosted instead of bought in

Error tracking runs on a self-hosted GlitchTip instance on our own Scaleway infrastructure, so crash reports never leave our servers. And they don't contain you: before a report goes out, user identity, IP, cookies, headers, authorization tokens, request body and query string are stripped. The body is the part that matters, because note content and search queries would otherwise sit in it. Performance tracing is off by default. This website's visitor stats are cookieless (Plausible, EU), and the app itself has none at all.

Keys that can do one thing only

Access keys follow the principle of least privilege: every system gets only the permissions it actually needs. Example: this website's automated deploy holds no account-wide CDN key, just a token that can do exactly one thing: purge this site's cache. The central keys themselves live in Scaleway's Secret Manager, not in external build systems.

The separation lives in the database, not in the code

In most services, application code is what keeps you from seeing someone else's data. One forgotten "where user = …" in a single query is then enough for a leak. In Memexa the separation sits a layer lower: PostgreSQL enforces it through row-level security. With no user context set, the database returns no rows: not the wrong ones, none at all. The context is set per transaction and falls away when it ends, so it can't carry over to the next connection out of the pool. To keep it that way, a check runs with the test suite that fires as soon as a new table arrives without row-level security or an explicit classification: the build fails instead of quietly passing.

What your AI may do is your call, and you can take it back

Claude, ChatGPT and Cursor don't get a master key. Access runs through OAuth: before you approve, you see which AI app is asking and where the access goes, and you confirm it yourself. The browser clipper and full AI access hold separate permissions: the clipper may file things, not search your store. You can disconnect AI apps any time in Memexa; an AI app can also hand its access back on its own. And access control doesn't hang on one central gate: every endpoint checks the session itself, so a mistake in one place doesn't open everything.

Deletion means deletion: in this order

You can delete your account yourself, confirmed by email. What happens then is deliberately the other way round: first every file in object storage is removed, across the complete user prefix, so orphaned files that no longer belong to any entry go too. Only then do the database rows fall. If the file deletion fails, the whole run aborts and your account stays, rather than leaving you with the records gone and the documents still sitting there. The run is safe to repeat. Individual items you can delete any time without touching the account.

The wires in between

Every leg is encrypted: TLS 1.2 and up at the edge, older versions switched off; database connections check certificate and hostname too, so they are not just “encrypted somehow”. The database sits on a private network with no public address. In front of it are a WAF and DDoS protection, and the origin server only accepts requests from the CDN's network, so you can't reach it directly. The application container runs without root, dependencies build from a pinned lockfile, and the one external binary we need is pinned to a version and a checksum.

What we do not claim

A page that only lists strengths is an ad. So here's the other list. Every item is one we've been asked about, or would be:

  • "Your data never leaves the EU."

    We don't say that. The moment you connect Claude or ChatGPT, you send sources to their provider. That's the point of the thing, but it's your decision, not our guarantee. Add the payment chain, named openly above.

  • "Only you can decrypt your data."

    We don't say that. Memexa is not end-to-end encrypted. For text recognition, search and linking to work, the service and the named processors have to be able to read the content. If you need genuine zero-knowledge storage, you need a different tool, and you should distrust anyone promising both at once.

  • "Certified to ISO 27001 or SOC 2."

    We're not. We've reviewed against OWASP-oriented criteria, which is a different thing from a certificate, and we don't call it one.

  • "Independently penetration tested."

    No. There are several documented, multi-pass security reviews of the code, all of them internal. An external pentest is a different thing, and until there is one, no other word goes here.

  • "Highly available, 99.9%."

    No. The production database runs without failover. There are backups and point-in-time recovery, but no committed availability level. And as long as we can't show you a rehearsed restore with a date on it, we won't promise one.

Documents & reporting channels

Memexa is bootstrapped: no investors, no exit pressure. Your subscriptions are the business model, not your data.

Try free for 14 days

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

See pricing