Field notes.
Notes from the build. Research we keep returning to, problems we are working through, and the long-horizon view of EU-sovereign AI memory.
-
Sovereign Inference Is Not Sovereign Memory
Mistral Medium 3.5 runs on 4 GPUs. That solves inference — not where your knowledge lives, who controls it, or how to prove it under the EU AI Act.
Read post -
Beyond Retrieval: World Models for Knowledge Memory
Retrieval finds stored items. Memory generalises from experience. The gap between them defines the next research problem in AI knowledge systems.
Read post -
Privacy in RAG Is an Architecture Problem
PRAG proves E2E-encrypted RAG is technically possible. But it only protects embedding vectors, not your documents.
Read post -
Retrieval Is Not Memory
Agent platforms call retrieval 'memory.' A new paper argues this is a category error — and the architecture choices get harder once you take it seriously.
Read post -
Graph Structure Is Not an Audit Trail
A knowledge graph makes GraphRAG answers better, not auditable. XGRAG shows what auditability requires, and why the architecture sits outside the model.
Read post