| Inspectable declarations | Collate publishes its entity model as open JSON Schema and documents its standards crosswalk. Ontology Studio and the Ontology Explorer make the graph visible to people. | The memory is plain Turtle in a git repository, so every declaration can be read, grepped and diffed on GitHub. Nothing is hidden in a database. |
| Validated declarations | The article says metadata is checked with SHACL shapes, covering cardinality, datatypes, identifier formats and required ownership. | The store has SHACL shapes for session document entities, index list items and HowTo steps, checked with pyshacl. The shapes are report-only: at build time 231 of 320 session files conform. Procedural gates also run. check-memory-gate.py, a Claude Code PreToolUse hook, blocks writes until the core files are read. validate-memory-protocol.py audits Claude Code JSONL transcripts. session-graph-gate.py, which is harness-neutral, compares triple sets with the named graphs in the triple or quad store the memory is loaded into (Virtuoso in this deployment). |
| Exportable declarations | The article describes Turtle, JSON-LD 1.1 contexts for each domain, and a documented API for export. | Turtle is the default and only serialization. AGENTS.md uses another format only when the user names it, so export means copying files or loading them into a SPARQL store. |
| W3C standards alignment | The article says glossaries are SKOS, relationships are OWL object properties, and lineage uses PROV-O, with SHACL for validation. | RDF and Turtle with schema.org vocabulary and a local onto: ontology. It now uses W3C SHACL for structural shapes, but there is no SKOS scheme or OWL axiom set for the memory itself; the corpus term registry uses owl:equivalentClass. |
| Provenance and history | Every entity is versioned, and each change to a description, owner or classification is emitted as an event on a timeline. | Each session file is a per-session record: its name carries the date, model and environment, and schema:dateCreated and schema:dateModified say when it was written. The store keeps no event log of its own: each session file is a snapshot record for one agent run, and the harness is generic, so any agent can write one. The Claude Code hooks and the transcript audit are specific to one harness; other harnesses have no transcript audit. Routing traces belong to the separate llm-routing-skill, not to the memory store. Repository history comes from git. |
| Lineage and graph structure | Lineage edges across sources, pipelines and dashboards, with transitive inference rules. Classifications propagate along lineage. | No data lineage. The graph links sessions to HowTo files through rdfs:seeAlso and onto: relations, which is a navigation graph, not a data-flow graph. |
| Agent access path | An MCP server, an SDK, and natural-language chat over governed context, with SPARQL reachable through tools. | Agents read files directly or query a SPARQL endpoint chosen at session start (AGENTS.md Step 0). There is no MCP memory server, by design. |
| Sovereignty and lock-in | The article's claim: open JSON Schema and W3C standards mean context is not locked in the platform. It is an assertion about design, not an independent audit. | The harness is files and rules: Turtle graphs in git, read either from the filesystem or through SPARQL against whichever endpoint a session selects (AGENTS.md Step 0). A move from one triple or quad store to another reloads the same Turtle and changes the endpoint; the rules and the harness stay the same. The platform-coupled part is the sync step in session-graph-gate.py, which emits Virtuoso isql or Graph Store PUT. A different store needs a new sync emitter, not new rules. Private overlays stay local, and credentials are referenced by Keychain (or host OS equivalent) item name, not stored in the Turtle. |
| Governance and AI disclosure | An audit log of every AI action, a badge on AI-generated content, and a mapping to EU AI Act obligations including Article 50, per the article. | Behavioral rules live in AGENTS.md and preferences.ttl. Actions are auditable through session transcripts, but there is no product-level audit log or AI-content labeling. |
| Scale and deployment | The article reports more than 700 JSON Schemas, more than 130 connectors, and enterprise deployments including a reported OpenAI internal data agent. These are vendor-reported figures. | Scale comes from the deployment platform, not from the harness. Today the memory is about 190 HowTo files and about 320 session files, held as files in git. The same Turtle loaded into an enterprise-grade triple or quad store inherits that store's scale. The design goal is loose coupling: rules and files any application or platform can read, not tight coupling to one application's schema or one vendor's API. |