Most people use AI for learning the shallow way: ask a question, get an answer, forget it, ask a similar question next week. That's a search engine with better manners. A tutor is different — it has a plan, it knows where you are in it, and it pushes you toward a goal. An agent with persistent memory can be that, and it's one of the more rewarding personal use cases because the payoff compounds the longer you stick with it.
What a tutor agent does that a chatbot doesn't
Four things. It sequences — turns "learn Rust" into an ordered path from ownership to lifetimes to async. It quizzes — tests recall instead of just explaining. It remembers — knows you keep missing the same concept and circles back. And it adapts — speeds up where you're solid, slows down where you're not. The first two a chatbot can fake in a single session; the last two require state that survives between sessions, which is what makes this an agent use case rather than a prompt.
How to set it up
1. State the goal and your level. "Get me to conversational Spanish in three months, I know maybe 200 words." The agent uses this to scope the course.
2. Let it build the curriculum. The agent outlines lessons, generates exercises, and prepares quizzes. On Hermes this becomes a skill that refines its teaching of you specifically over time; Claude Code or Codex are better if the subject is itself coding and you want it to run and check your exercises.
3. Save progress as state. After each session the agent records what you covered, what you missed, and where you are. This is the part that makes it a tutor and not a generator.
4. Optional — schedule reviews. If the agent runs on a host, it can send spaced-repetition prompts at the intervals that beat forgetting, rather than waiting for you to remember to study.
When to deploy it vs run on demand
Tutoring is interactive, so unlike the finance auditor it doesn't have to run unattended. Run it on demand if you only study when you sit down with it. Deploy it on a server if you want it to keep state reliably and nudge you to practice on a schedule — the same persistence that powers the other use cases in the overview, applied to learning.
If you'd rather run your tutor agent on your own machine and pick up lessons from your phone between sessions, Onepilot lets you deploy and reach Hermes, OpenClaw, Claude Code, or Codex on a remote host without wiring up the connection yourself.