A new academic framework wants scientists' private notes to become shared, searchable knowledge instead of dying in someone's personal files.
Researchers describe Valhalla, a layered knowledge-state and service-governance framework built for long-term scientific work with LLM agents. Instead of the flat graph of nodes and relations most tools use, Valhalla runs a five-layer File-Resource-Entity-Relationship-Graph model that separates source documents and provenance from extracted concepts, semantic judgments, and task-specific views. A second piece, a Router-Contract-Workflow architecture borrowed from microkernel operating-system design, governs how language models can read, edit, or extend that knowledge without breaking its structure. The team tested a prototype on an antibody-design literature review covering 26 papers, 80 knowledge entities, and 92 semantic relationships.
Most LLM memory and knowledge-graph tools are built around one person's workflow, so a lab's collective understanding ends up scattered across notes that don't talk to each other. By treating provenance and semantic judgments as separately governed layers instead of mixed-in graph nodes, Valhalla is going after the harder problem of merging different researchers' knowledge into one auditable structure, not just retrieving facts faster.
It's still a single case study, though, and the real test is whether other labs bother wiring their notes into someone else's five-layer schema.