Agentic concepts, explained

Hermiq is built out of a handful of ideas that come from the world of AI agents. If you have never met these words before — agent, skill, memory, context, tool, MCP, RAG — this page explains all of them in plain language, in the order they make sense. Nothing here assumes you have used an AI system before.

Read this page once, and the rest of the documentation will make sense.

Start with the problem

A chatbot answers one question at a time. You ask, it replies, and it forgets. That is fine for a quick question, but it cannot do a job for you.

What most people actually want is something closer to a colleague: someone you brief once, who knows your organisation's material, who can look things up and take action, who remembers what happened last week, and who does the job every Monday morning without being asked.

That is what an agent is for. Everything below is one piece of turning a forgetful chatbot into something that can hold a job.

The pieces, in one table

Concept	In one sentence	Everyday analogy
Agent	The worker: a configured AI with a job description.	The colleague you hired.
Skill	A reusable instruction sheet teaching a way of working.	A how-to guide in their desk drawer.
Memory	What the agent remembers between conversations.	What they learned about you over months.
Context	Reference material you hand the agent for this job.	The project binder on their desk.
Tools & MCP	The actions the agent can actually take.	Their hands, phone, and keyboard.
RAG	Looking things up before answering.	Checking the filing cabinet before speaking.
Runs & schedules	The agent doing its job, once or on a rhythm.	A shift, or a standing Monday task.

Two more pieces exist to keep all of this safe, and they matter as much as the rest: approvals (a human says yes before something risky happens) and incidents (a written record when something goes wrong).

And underneath all of it sits one architectural choice — the safe setup. The model runs in a hardened container on your own hardware with no outside access at all: no internet, no files, no tools. It only thinks. Every file, lookup and action goes through Hermiq instead, which does it as you and only if you granted it.

Which is what lets you answer the question everyone actually has: "can I have an assistant that summarises my mailbox every morning, but can't read my files or delete my mail?" Yes — and not because you asked it nicely in a prompt. That page has the diagram.

How they fit together

Think of hiring someone:

1. You hire them → that is the agent. You give it a job description (a prompt) and decide which AI model powers it.
2. You give them the house rules → those are skills. "This is how we write a permit decision." Reusable, and shared across agents.
3. You hand them the project binder → that is context. "Here is the design document for this project, our standards, who the stakeholders are."
4. You show them where the filing cabinet is → that is RAG. They look things up in your files and records instead of guessing.
5. You give them a phone and a keyboard → those are tools, offered through MCP. Now they can actually do things, not just talk.
6. They remember you → that is memory. Next month they still know your preferences.
7. They work Monday mornings → that is a schedule, and each time they work is a run.
8. They ask before anything drastic → that is an approval.

Every one of these is optional. An agent with none of them is just a chatbot. Add them as the job requires.

The one distinction people trip over

Skill vs context sounds like the same thing — both are text you give the agent. The difference is reusability:

• A skill is a way of working that travels. "How to summarise a meeting" is useful to any agent, in any project, at any organisation. Skills can be shared and published.
• A context is this situation's material. Your project's design.md, your team's naming conventions, this quarter's targets. It is specific to you and it does not travel.

Rule of thumb: if you would happily give it to another organisation, it is a skill. If it only makes sense inside your project, it is context.

Memory vs context is the other one:

• Context is what you hand the agent, deliberately.
• Memory is what the agent picked up by working, without being told.

What makes Hermiq's version different

Hermiq runs all of this inside your own Nextcloud. That matters for reasons that are not about AI at all:

• Your data stays yours. The agent reads your files and records where they already live. Nothing is shipped to someone else's platform to be useful.
• Everything is a record. Agents, their skills, memory and context are all stored as ordinary OpenRegister objects — which means every change is versioned and audited like any other data in your organisation.
• It is governed by default. Because every run and every tool call is written to an audit trail, and risky actions can be gated behind a human approval, an agent doing real work is something you can actually account for — including under the EU AI Act.

That is the whole idea: an agent you can hand real work to, that behaves like part of your organisation rather than a service outside it.

Where to go next

• New to all of this? Read Agents next — everything else hangs off it.
• Want the agent to know your material? Context and RAG.
• Want the agent to do things? Tools & MCP.
• Worried about your data leaving? The safe setup — start there before you configure anything.
• Responsible for governance? Approvals and Incidents.