Memory you can audit: citations, retraction and GDPR-grade privacy

There's a vast difference between an AI that *sounds* confident and one that can *prove* what it says. For any production system touching real user data, that difference is the line between a feature and a legal liability.

Citations: provenance isn't optional

When MemoryFirst answers, it doesn't hand you a free-floating claim. It hands you the claim and its source: the paragraph in the document, the minute in the recording, the exact message in the thread. This changes three things for a developer:

• You debug retrieval, you don't guess. If an answer is bad, you see *which* passage caused it.
• Your user trusts it. "Per the contract, clause 4" is verifiable. "I think so" is not.
• You stay compliant. Tracing a claim back to its origin is the foundation of any serious audit.

Retraction: the right to be forgotten, for real

The GDPR gives people the right to have their data deleted. In a memory layer, "delete" can't just mean dropping a row: derived knowledge has to stop showing up in answers. MemoryFirst treats retraction as a first-class tombstone: a retracted source is excluded from retrieval, and stops contaminating what the AI says.

EU data, and never for training

Two lines we don't negotiate:

1. EU residency. Infrastructure on European soil (Hetzner). Self-hostable in your own cloud if you need it.
2. We never train models on your data. Your knowledge is yours. It feeds neither our models nor third parties'.

An AI that asserts without citing the source isn't memory: it's a hallucination with a good reputation.

Why this is an edge, not a burden

It's tempting to see privacy as a compliance cost. For a product selling into regulated sectors, it's the opposite: it's the pitch. When you can say "EU data, every answer cited, real retraction, no training," you stop competing on features and start competing on trust. And trust isn't replicated by a bigger model.

The technical details are on the security page.