“EU hosting” now appears on almost every page that sells a memory for your AI app. It usually means where the database sits. But an AI memory does more with your data than store it: it reads documents, computes embeddings for search and has a language model pull out facts. Each of those steps sends text to a model provider, and that provider can sit somewhere other than the database.

So this comparison separates two questions that otherwise vanish into one column: Where is the data stored? And: Where does the AI processing run? Six providers are compared. One of them is our own product, Memexa; the other five we filled in strictly from their own pages, with a source for every entry.

Storage location is not processing location

In an AI memory, your content passes through three layers:

  1. Storage. Database, files, backups. This is the layer “EU hosting” almost always refers to.
  2. Processing inside the memory. OCR for scans, embeddings for semantic search, a language model that condenses notes or extracts facts. Here text goes to a model provider, often OpenAI or Anthropic in the US.
  3. The AI app you talk to. Claude, ChatGPT or Mistral Vibe receive the citations the memory returns. What happens there is governed by your contract with the AI app’s provider, not with the memory.

So a provider can honestly write “servers in Germany” and still send every note through a US model for search. StudioMeyer says as much on its own product page: EU database hosting does not mean exclusively European processing. That’s a fair statement. The question is whether you find it before you sign.

No memory solves the third layer for you. If everything is to stay in the EU, you also need an AI app whose processing runs in the EU.

The comparison

Where data is stored and where the AI runs

As of November 3, 2026, from the providers’ own pages (sources at the end). “Not stated” means we didn’t find it there, not that it doesn’t exist.
ProviderWhere data is storedWhere AI processing runs
MemexaScaleway, Paris (France)Mistral, Paris: OCR, embeddings, condensing. Re-ranking of search hits at IONOS, Berlin
Cognee CloudAWS us-east-1, USA. EU only self-hosted or as an Enterprise deployment in your own EU cloudModel gpt-oss-120b; who runs it and where is not stated
Mem0 PlatformNot stated. The privacy policy names the US and other countriesAI providers “such as OpenAI and Anthropic”, region not stated
SupermemoryNot stated (“various geographic locations”). Strict data residency only in the Enterprise offeringOpenAI and Google Gemini, region not stated
StudioMeyer MemoryHetzner, Nuremberg (Germany)OpenAI for embeddings, Anthropic for AI features; per the provider, not exclusively in Europe
Basic Memory CloudUSA (Fly.io, Neon, Tigris)Agents run on a DeepSeek model at Fireworks AI, region not stated

Contract, export and deletion

As of November 3, 2026, from the providers’ own pages.
ProviderDPASubprocessorsExportDeletion
MemexaOn requestPublic list with seat and place of processingMarkdown, JSON and original files as ZIP, even without a subscriptionDelete your account yourself, confirmed by email
Cognee CloudYes (Art. 28 GDPR), not linked publiclyPlatform services in the privacy notice; list for content only in the DPANo export documented in the API referenceDatasets via UI and API; content 30 days after the contract ends
Mem0 PlatformFor business customers, not publicNo public list found (trust center not checked)JSON via API, in a schema you defineAccount deletion, runs in the background; no fixed period stated
SupermemoryOn request or in the Enterprise offeringNo public list; on request in the “legal pack”JSON from settings, up to 25,000 itemsDelete your account yourself
StudioMeyer MemoryNot found for MemoryNo list; named are Hetzner, OpenAI, Anthropic, Resend, SentryJSONAutomatically 30 days after the paid term ends
Basic Memory CloudNot foundProviders in the privacy policy, without countriesMarkdownWithin 30 days, backups up to 90 days

The providers one by one

Memexa

Memexa runs on Scaleway in the Paris region: servers, database and document storage. OCR, embeddings, condensing and audio transcription run on models from Mistral AI in Paris. When a search returns several similarly good hits, an additional model sorts them, operated by IONOS in its Berlin data centre; it receives text excerpts, not whole documents. No US model provider sees your content, and nobody trains models on it. Every processor is listed with seat and place of processing on the trust page, including those outside the EU (the payment chain works with US subcontractors but never receives content).

What Memexa doesn’t have: no certificate, no end-to-end encryption, no external pentest. You get the DPA on request; it isn’t linked publicly. And the third layer above applies here too: if you ask in Claude or ChatGPT, the excerpts found end up with Anthropic or OpenAI.

Cognee

Cognee has its own page on EU hosting, and AI answers cite it often. It describes two routes: run the open-source version yourself in any EU region, or a managed Enterprise deployment in the customer’s own EU cloud. For the hosted Cognee Cloud, however, the privacy notice says: hosted in the USA, AWS region us-east-1, no EU region offered to customers today. By its own account, Cognee claims no SOC 2, HIPAA or ISO certification. If you want Cognee in the EU, you run it yourself or negotiate Enterprise.

Mem0

Mem0 is headquartered in the US, and its privacy policy names the US and other countries as places of processing. We found no EU region on the pricing page, in the docs or in the terms; on-prem deployment is part of the Custom plan. As AI providers, Mem0 names OpenAI and Anthropic. The DPA is available to business customers. Export runs via API as JSON, account deletion in the background.

Supermemory

Supermemory lists SOC 2 as certified on its security page. It doesn’t state where data is stored: the terms speak of various geographic locations. Stricter data residency is part of the Enterprise offering, business self-hosting starts with the Scale plan at $399 a month; a local version is free up to 10,000 documents. For AI features, the privacy policy names OpenAI and Google Gemini. DPA and subprocessor list are available on request from support.

StudioMeyer Memory

StudioMeyer, a provider based in Spain, runs its server and database at Hetzner in Nuremberg. OpenAI computes the embeddings, AI features run on Anthropic, and the product page says so openly. Prices are in euros: free, Pro at €9 and Team at €19 a month. We found no DPA for the Memory product.

Basic Memory Cloud

Basic Memory is first of all a local open-source tool: notes live as Markdown files on your machine. The hosted Basic Memory Cloud runs on servers in the US according to its privacy policy; the providers are listed there by name but without country. The Cloud’s agents run on a DeepSeek model operated by Fireworks AI. No DPA is mentioned. If you use Basic Memory locally, you’ve solved the storage question, but not where the connected AI app runs.

Five questions for any provider

The tables will age. The questions won’t:

  1. Where are the database, files and backups? Country and provider, not just “EU”.
  2. Which models process my content, and where? OCR, embeddings, extraction, re-ranking: each step separately.
  3. Is there a DPA, and is the subprocessor list public? With seat and place of processing, not just company names.
  4. Can I get everything back out? In which format, with original files, and after the subscription ends?
  5. What does deletion mean? Immediately or after a period, and how long do backups last?

A provider that won’t answer question 2 has usually answered it already.

Frequently asked questions

Is an EU-hosted AI memory automatically GDPR compliant?

No. Storage location is one part of the answer. The others are where AI processing sends text, whether there’s a DPA, and how export and deletion work. Whether your specific use is permitted also depends on which data you put in. None of this is legal advice.

Can I run a memory myself in the EU?

Yes. Basic Memory runs locally, Cognee and Supermemory can be self-hosted, and Mem0 exists as an open-source library. Then storage sits with you. Embeddings and the language model run wherever you configure them, though: without your own models, that is often a US provider again.

What does an EU memory bring if I ask in Claude or ChatGPT?

It keeps the archive in the EU: your documents, scans and notes, and everything computed from them for search. The AI app only receives the excerpts that match a question, not your whole archive. To stay fully in the EU, use an AI app that processes in the EU itself. Which memory servers exist is compared in Memory MCP servers compared; how to connect Memexa is on the setup page.

Sources

All entries about other providers come from their own pages, retrieved November 3, 2026: