Core Primitives¶
Methodology → building blocks
The smallest objects that make memory auditable.
A durable memory system must represent reality in a way that survives model churn. That means: the unit of memory is not “a paragraph” — it is a structured object with provenance.
What we treat as first-class
Entities (subjects/objects)
Who/what exists in the domain: people, organizations, devices, molecules, policies, contracts.
Processes (events/transformations)
What happens over time: treatments, approvals, payments, failures, audits, experiments.
Relations (edges)
How entities connect: causes, depends-on, contraindicates, authorizes, violates, explains.
Sources (provenance objects)
Where we learned it: document IDs, versions, timestamps, authors, datasets, and extraction method.
Diagram: primitive schema (minimal)
flowchart TB;
S["Source"] --> F["Fact (claim)"];
F --> E1["Entity"];
F --> P["Process / event"];
E1 --> R["Relation (edge)"];
R --> E2["Entity"];Why provenance matters
Without provenance, you can’t audit, debug, or falsify.
Provenance turns “an answer” into a decision-grade artifact: it makes claims inspectable and contestable.
Audit
Which source supports this claim? Which version? Under what scope?
Debug
Is the failure data, extraction, constraints, or model behavior?
Confidence
Confidence attaches to evidence quality, not model fluency.
Falsification
What new data would change the conclusion? Which edge is uncertain?
Diagram: provenance chain (what “grounded” really means)
flowchart LR;
D["Document / dataset"] --> V["Version + timestamp"];
V --> X["Extraction"];
X --> C["Claim (fact)"];
C --> T["Trace step"];
T --> O["Output"];