Your AI agent wakes up a stranger every morning.

Agent RDF Memory: the folder that teaches it once, and makes it stick.

Presented by Claude Sonnet 5.5 via Claude Code, on behalf of Kingsley Uyi Idehen.

An overview of agent-rdf-memory/README.md01 / 10
The problem
“Use this output folder.” “Never guess a web address.” “Ask before you delete.” SESSION ENDS Next morning:“Who are you again?”
Fig. 1. Rules stated in chat last as long as the chat does.

You explain the same rules every morning, because chat is no place to keep them.

Each rule you teach is obeyed for one session, then lost. The more you care about how your agent behaves, the more time you spend repeating yourself.

Agent RDF Memory02 / 10
The big idea
# preferences.ttl (one rule, as data) :step-outputDirs a schema:HowToStep ; schema:position 1 ; schema:name "Route artifacts to model-specific output directories"@en ; rdfs:seeAlso <howto/artifact-routing.ttl> .
Fig. 2. A rule becomes a typed, linked record that any agent can read and any person can review.

Write the rules down as data your agent can query.

A queryable behavioral contract for AI agents, encoded in RDF-Turtle. The memory files, not the chat, are the authoritative source.

Agent RDF Memory03 / 10
How it works
1List thefolders 2Readcore.ttl 3Readpreferences 4Privateoverlay 5Readontology 6Readindex 7Classifythe task 8Followthe links 9Fall backto files Then, and only then, it acts.
Fig. 3. The retrieval protocol, as the README lays it out.

Nine steps stand between a cold start and an agent that knows you.

Files establish the baseline. A configured SPARQL endpoint, when there is one, selects the extra context the task needs. Either way, the agent has read its contract before it writes a word.

Agent RDF Memory04 / 10
The structure
Behaviorguide NINE THEMES Identity & WebIDMemory managementArtifact routingRDF authoringHTML & KG explorer Ontology generationSkill workflowsVirtuoso SPARQLTerminology Each step is a short pointer. The companion file holds the reasons.
Fig. 4. One hub, nine themes. Each rule links to a companion specification with its rationale, examples and gates.

312 rules. 183 specifications. 308 recorded sessions. Every rule with its reasons on file.

Counts taken live from the repository. The README's own table lists an earlier total of 105 steps, so the folder has grown well beyond it.

Agent RDF Memory05 / 10
Public and private
preferences.private.ttl Local only. Gitignored. Overrides the public defaults. 🔒 preferences.ttl Public-safe. Reusable rules and shared structure. wins
Fig. 5. The private overlay is optional, and the public file never depends on it.

Share the rules. Keep the secrets.

Endpoint order, identity defaults, private paths and credentials stay in a local-only overlay. A team can publish the harness without publishing anyone's personal setup.

Agent RDF Memory06 / 10
Smart retrieval
Your request Prompt intentontology.ttl SPARQL SELECTtopics, how-tos, sessions Only the rules that matter Direct file reads Endpoint offline? Read the files.
Fig. 6. Ontology-routed context selection, with a file-based fallback.

A good agent doesn't read everything. It reads what this task needs.

The ontology classifies the request and routes it to relevant topics and how-tos. The lookup runs against any configured RDF store, and degrades gracefully to plain files.

Agent RDF Memory07 / 10
Growing the memory
1Choose theowning theme 2Add a HowToSteppreferences.ttl 3List it under thetheme's steps 4Write the full spechowto/<topic>.ttl 5New theme? Updatethe README Say it once. It is remembered.
Fig. 7. The README's five-step recipe for adding a rule.

Teaching your agent something new takes five steps, not five reminders.

A concise pointer in the hub, a full specification beside it. The rationale, the gates and the story of why the rule exists all live in one place.

Agent RDF Memory08 / 10
Portable by design
Turtle filesthe source of truth An endpoint + a graphhttps://{CNAME}/sparql Virtuoso (the example) Any RDF platform Yours, next
Fig. 8. {CNAME} is a placeholder for whichever endpoint you select. It is never a hard-coded host.

It needs an endpoint and a graph. Not a vendor.

The README's worked examples use Virtuoso, including local and RDF Import DET setups, but the pattern is platform-neutral. And because the memory is Turtle, it is portable, diffable and yours.

Agent RDF Memory09 / 10
Start today

An agent that remembers.

1. Copy agent-rdf-memory/ into your project 2. Fill in core.ttl with who you are 3. Point your agent at it, then tell it once

Source: agent-rdf-memory/README.md in the ai-agent-skills repository from OpenLink Software

Generated by Claude Sonnet 5.5 via Claude Code, on behalf of Kingsley Uyi Idehen. Narration by Leo, xAI Grok text-to-speech. Linked Data resolved via URIBurner (Virtuoso-backed).

An overview of agent-rdf-memory/README.md10 / 10