Ontofelia

The AI agent gateway with a semantic soul.

An open-source, self-hosted agent gateway that gives your AI a persistent identity, structured long-term memory powered by OWL ontologies, and secure multi-channel communication.

Getting Started · Architecture · Documentation · Contributing

---

⚠️ v0.0.1 — early research preview

Ontofelia is an active research project in neuro-symbolic AI: it studies whether coupling an LLM with an OWL reasoner and a governed RDF knowledge graph yields a more reliable agent than an LLM with conventional string/vector memory. It is under heavy development and not production-ready — APIs, behaviour, and results will change.

Built by AI — a "vibe coding" experiment. Ontofelia was implemented end-to-end with AI coding agents — Claude Code, Codex, and Gemini — under human direction. It is as much an experiment in AI-assisted, vibe-coded development as it is a neuro-symbolic AI study. Expect the rough edges, inconsistencies, and AI-generated artifacts that come with that.

On competitor comparisons — please read. We have not run a validated, controlled head-to-head against the real Hermes or OpenClaw products, and we make no claim to have beaten them. The evaluation below is against a fair RAG baseline inside our own harness (a research proxy for string/vector memory) — not the shipped competitor binaries. An exploratory real-agent run we attempted surfaced harness and scoring problems (and a bug in our own forgetting path, since fixed), so it is not a published result. Treat every number here as preliminary research, not a verdict.

---

Why Ontofelia?

Most AI agents are stateless parrots — they forget everything between sessions, can't reason about what they know, and store context as opaque embeddings. Ontofelia is different.

Feature	Typical Agent	Ontofelia
Memory	Vector embeddings / flat text	RDF triples + OWL ontology
Reasoning	None	OWL inference (via Oxigraph + Reasonable)
Knowledge query	Fuzzy similarity search	SPARQL 1.1 — precise, structured
Self-awareness	None	Ontology reflection — the agent can inspect and explain its own knowledge model
Channels	Single platform	Multi-channel — WebChat, Telegram, Discord, WebSocket nodes
Tools	Basic function calling	Governed execution — policy engine, Guardian approval, audit log
Extensibility	Hardcoded	Skills + Plugins — hot-loadable ESM modules

Ontofelia doesn't just remember — it understands.

The research question

Most agents — including OpenClaw and Hermes — store long-term memory as strings or vectors that don't reason. Ontofelia's research hypothesis is that a governed, reasoning knowledge graph can offer truth-maintenance a retrieval index cannot guarantee:

Hypothesis (under test): an agent whose memory reasons can forget completely when told to, and reliably flag contradictions — where a retrieval-augmented agent is unreliable.

This is a hypothesis we are testing, not a settled claim, and explicitly not a head-to-head result against the OpenClaw or Hermes products. On plain recall a good vector-RAG already matches us; the open question is the part a retrieval index structurally cannot offer — deletion you can trust, contradictions caught, and beliefs you can audit (memory_explain — "why do I believe this?").

📊 Evaluation — honest & reproducible

Scope: this is a controlled research benchmark against a RAG baseline — a fair, strong vector-RAG that stands in for the string/vector memory style of agents like OpenClaw and Hermes. It is not a benchmark of those products themselves. The harness swaps only the memory behind one LLM — Ontofelia's governed knowledge graph vs the RAG baseline vs no memory — across 3 models, scored H0–H6 with McNemar/Wilcoxon significance (Holm-corrected). Pilot run pilot-2026-06-25:

Capability	Ontofelia	fair vector-RAG	Verdict
Recall (H1)	0.99	1.00	tie — we don't claim a win
Multi-hop (H2)	1.00	0.94	~tie (not significant)
Forget on command (H6)	1.00 — 0% leak	0.54 — leaks the "forgotten" fact 31–62%	decisive win, every model (p<0.001)
Contradiction flagging (H3)	0.93	0.20	decisive win, every model (p<0.001)
Auditable provenance (H4)	0.89	0.64	win (pooled, p=0.014)
Constraint/consistency (H5)	1.00	0.89	modest

The two effects that held on every model tested were forgetting (H6) and contradiction detection (H3). This supports the hypothesis against a RAG baseline — it is not evidence about the OpenClaw or Hermes products, which we have not benchmarked. Pilot-scale, preliminary.

Run it yourself:

This is the beginning of the README. Read the rest at the original. Read the full README ↗