Video: "Hermes Oracle Agent Just Changed AI FOREVER!" by Julian Goldie on YouTube.
What Oracle actually is
Oracle is a voice-control layer built on top of Hermes Agent OS. You set a wake word — Julian uses "Jarvis" — and from that point the agent listens for spoken instructions. Ask it to build a flashcard app, generate an SEO landing page, or pull up recent data, and it sets to work. You watch the code appear and the output load; you do not type a single prompt.
The underlying model can be anything. Julian runs it on the Free Oracle — a local model from Nous Research that is wired directly into Hermes, costs nothing to use, and does not phone home to a cloud API. That matters if you are handling client work or anything commercially sensitive.
What the demo actually shows
Julian ran Oracle through several tasks in one session. He built a working Japanese flashcard game by voice, asked Oracle to generate an SEO website structure, had it open a browser to Bing and fetch current figures, and then asked it to teach him three Japanese phrases — which it did, complete with spoken examples. Each task took seconds from instruction to output.
The framing throughout is that this is a Jarvis-style assistant, and in practice that comparison is reasonably accurate. The agent is not just answering questions; it is taking actions, generating files, and displaying results — all from a spoken prompt. That is a meaningful step up from the earlier voice demos, which were largely question-and-answer flows.
The memory piece
Every output Oracle produces gets saved into what Julian calls a "memory galaxy" — an Obsidian knowledge graph that visualises connections between saved notes, sessions, and outputs. In practice this means the agent accumulates context over time: it knows what you built last Tuesday, what research it pulled two weeks ago, and what decisions you made during earlier sessions.
That persistent memory is what distinguishes Oracle from a smart dictation tool. An agent that forgets everything after each session is useful for one-off tasks. An agent that carries context forward becomes something closer to a member of staff who gets better the longer you work with them.
What is genuinely new here — and what is not
We already covered the Hermes voice setup with Minimax M3 in June, and Julian showed a Jarvis build in the same week. So the voice mode itself is not new. What Oracle adds is the combination: keyword wake, app generation, active browser control, and memory logging all in a single session, running locally on a free model.
Worth knowing: the demo is clean because Julian is an experienced operator who has been running Hermes for months. First-time setups take longer. The wake-word detection is software-level rather than always-on hardware, so it listens only when you are at the machine. And the memory galaxy needs Obsidian installed and configured — it does not set itself up. That said, for anyone already running Hermes Agent OS, Oracle is a genuinely useful upgrade rather than a feature demo that lives in a separate tool.
Where this connects to NordSys
We install and configure Hermes Agent OS for UK businesses — including the memory layer, model selection, and voice setup. If Oracle is the kind of workflow you want — speaking instructions and watching your AI build things — we set it up properly so it actually works in your environment, not just in a demo. The difference between a clean YouTube demo and a business-ready agent is usually configuration, and that is exactly what we handle.
See our AI Agents service →