Structured memory substrates

The claimed advantages of an explicit knowledge layer are introspectability by humans and machines alike, enforceable access boundaries, and cheaper, more faithful answers than vector-only retrieval. The supporting numbers in this scan come from a single cited study — more than twice as many questions answered, at 80% better fidelity, with fewer hallucinations and lower token cost. Structure is not only a retrieval trick: one production example executes a rules graph across 15,000 cities in-graph, as procedural memory rather than as documents about procedure.

A second strand argues the same point from the warehouse side. What agent memory needs is versioning, temporality, concurrent multi-writer access, conflict resolution and entity resolution — not retrieval over a transcript pile. The entity-resolution examples are mundane and decisive: a pronoun resolving differently on different days, a canonical definition of revenue, a fiscal year that means one thing per tenant. The volume argument runs underneath it: agents produce far more data than humans, so the substrate has to be one that survives concurrent writers by construction.

A third strand generalizes past graphs entirely: whatever the corpus, compile it into a substrate the agent navigates natively — a virtual filesystem, a domain-specific knowledge representation — rather than making the agent query human-oriented APIs or stuff raw retrieved chunks into context. The reported result is pointed: a compiled virtual filesystem over papers and biological databases outperformed the same models given protocol-mediated access to the same corpus, roughly ten times faster and cheaper, because brittle human-oriented APIs were the bottleneck. The matching anti-pattern, named from the retrieval side, is a research agent running many searches and filling its window with raw chunks.

Also: GraphRAG; GQL and Cypher; the UK data-and-AI-innovation GraphRAG study; Gartner's "context layer as the central brain"; Cognee and its Claude Code plugin; PaperClip; MCP.

The defence is the substrate-compilation evidence. An agent-native file substrate beat protocol-mediated access to the same corpus, roughly ten times faster and cheaper — a direct data point in the files-versus-protocol debate, and a validation of Assay's markdown-everything plane: memory files, briefs, the derived board. Two consecutive spotlight companies betting against context-stuffing is the same instinct behind Assay's self-contained-brief rule, which exists so a worker never has to assemble its own context by search.

Plenty of questions here — which briefs touched this guard, which findings does this ruling supersede, which streams depend on this spec — are answered today by an agent re-reading files at token cost, where a structured index could answer them deterministically. statusgen already derives a board from the registers, which is a narrow instance of exactly this; the graph argument is that the derivation could be typed and general rather than board-shaped.

First, temporality and versioning on memory entries: a validity window plus a supersedes link, rather than entries that are simply true until someone notices they are not. Second, first-class conflict resolution for concurrent writers — several roles plus a consolidation pass all touching one memory plane, a hazard Assay already knows from its single-writer rules for generated files. And one design instruction for the unbuilt consolidation pass: make its output typed nodes and edges — belief, evidence, supersedes, derived-from — with citations as edges. That is a structural gain even if the storage stays markdown, because it gives the human accept-or-reject step something diffable instead of a paragraph.

• Whether the graph-versus-vector fidelity gap replicates outside vendor-adjacent studies; the headline figures come from a single cited study.
• Whether anyone reports a graph memory layer being maintained over years — the failure mode for ontologies is schema drift, and no source in this scan addressed it.
• Whether agent-native substrate compilation shows up outside science corpora, which would move it from a domain trick to a general pattern.
• Whether a small-scale rather than enterprise-scale deployment reports the graph paying for its maintenance cost.