FAQ AI/Hermes Questions

1. Can Hermes build strong MD files, or do humans have to do it?

Honest answer: both. Hermes can draft MD files and iterate on them, but the judgment calls, the strategic intent, the "this is what actually matters" layer, that has to come from you. Think of it as collaborative authoring. You supply the direction and the corrections, Hermes handles the formatting and maintenance. If you let it write the whole thing without supervision, you get plausible-looking files that don't actually reflect how you work.

2. Is hallucination management just correcting every error and logging it to the MD files?

Pretty much, yes, but the key is narrowing the scope of each agent's job. Hallucinations spike when an agent is asked to hold too much context and make too many judgment calls in one session. The 3-agent team structure Justin showed (narrow tasks, one manager) exists precisely for this reason. Smaller jobs per agent = fewer hallucination opportunities. Corrections still need to go into the MD files AND GBrain so they persist across sessions.

3. Should corrections also go into GBrain?

Yes. MD files are the agent's operating rules. GBrain is the long-term memory that survives session resets. For anything you want the agent to "always know," GBrain is the right place. MD files for behavior, GBrain for facts and history.

4. How to deal with Telegram bloat?

Justin handles this with the /fresh command, which you've already discovered. The tradeoff is real: long sessions = more context = better continuity, but also higher token cost and drift. The way to reduce drift without losing continuity is to keep the important decisions in GBrain and the MD files, not in the chat thread. The thread is working memory. GBrain is long-term memory. If the right facts are in GBrain, a fresh session can be almost as capable as a long one.

5. Does an agent ever need to be sacked and rebuilt from scratch?

Sometimes yes. The signal is when correcting one behavior breaks another, or when the agent has accumulated so many conflicting instructions that it's inconsistent no matter what. A clean rebuild with better-structured MD files and a well-seeded GBrain usually performs better than a heavily patched agent. The $1,500 AUD you spent isn't wasted though, you now know exactly what the right instructions are. That knowledge is what you put into the new build.

The short version: narrow the tasks, keep corrections in GBrain and MD files, use /fresh more aggressively, and don't be afraid to rebuild if the foundation is shaky.