RDF Memory Protocol

Step 0 — Load-Path Gate (before the file-read sequence)

Before executing the mandatory file-read sequence below, check the most recent session TTL (newest file in sessions/) for onto:usedFileReads or onto:usedSparqlEndpoint:

• Found → skip elicitation; load memory via the recorded method.

• Absent → STOP and elicit from the user:

  "Memory loading preference for this session:\n> (1) File reads [default]\n> (2) SPARQL — localhost:8890/sparql/ [local Virtuoso]\n> (3) SPARQL — other endpoint / other URI"

  Record the choice in the current session TTL as:

  • :session onto:usedFileReads true . (option 1)
  • :session onto:usedSparqlEndpoint <{endpoint-url}> . (option 2 or 3)

  Wait for the user's selection, then proceed to the load sequence below using the chosen method. On all subsequent sessions where a recorded choice exists, skip this elicitation silently and reuse the recorded method.

Mandatory Retrieval Sequence

Before responding to the first user request in every session, load the RDF memory from /Users/kidehen/Documents/Management/Development/ai-agent-skills/agent-rdf-memory/:

1. Read core.ttl for user identity and output-routing requirements.
2. Read preferences.ttl for the public operational-rule manifest.
3. If present, read preferences.private.ttl as a local-only private overlay; private overlay values take precedence over public defaults and must not be required for public reuse of the memory harness.
4. Read ontology.ttl for prompt-intent classes, retrieval policies, and ontology-routed context selection rules.
5. Read index.ttl for session pointers.
6. Determine the current prompt intent, then prefer SPARQL context selection against the configured endpoint to retrieve relevant preference topics, rdfs:seeAlso howto documents, and recent-session context.
7. Treat {CNAME} in SPARQL templates as the selected endpoint host or endpoint family, not as literal localhost. Determine the operating value from the private overlay when present, otherwise from public endpoint configuration or the current task context.
8. If SPARQL context selection is unavailable or incomplete, fall back to direct file reads and read the most recent file in sessions/ for previous-session lessons.
9. When a rule applies, follow its rdfs:seeAlso link to the referenced howto/*.ttl file before acting.

Treat these RDF files as authoritative operational memory. Do not substitute a summary, MEMORY.md, or assumptions for them. Re-read the relevant howto file when a task triggers one of its rules.

RDF Notation Preference

On the first RDF authoring task, check the private preference overlay and recent session memory for a notation preference. If none is recorded, offer Turtle (default), LD-Vault, YAML-LD, JSON-LD, and Type your own RDF notation preference. If the user does not answer and the task should proceed, use Turtle as the default. Record the selected preference locally in preferences.private.ttl when available, and in the current session record; never treat Turtle as a requirement. Honor a recorded preference, and update it whenever the user asks to change it. Read howto/rdf-serialization-default-clarification.ttl for the detailed rules.

The source design is documented in /Users/kidehen/Documents/Management/Development/ai-agent-skills/agent-rdf-memory/SESSION-START-HOOK.md. That document describes a Claude Code hook; this AGENTS.md is its Codex equivalent instruction entry point.