Configure Hermes Agent's project instructions, persistent personality, and memory—and build an agent that keeps getting better.
Most people meet an AI agent the way you'd meet a vending machine: walk up, type a request, get something back, walk away—and next time, start over. This book is about the alternative.
Picture your configuration actually finished: you close the editor, start the agent, and leave the room. Nothing in front of you is a chat window anymore—you've built an agent you configure once and improve continuously, instead of re-explaining it every session.
You ask your agent the same question on two different days—"The vendor missed the deadline again—what do I send them?" When Monday's answer comes back warm and apologetic and Wednesday's clipped and legalistic, you can tell the personality lives nowhere durable—and you know what to fix.
You restart the agent after an overnight release, ask it to pick up yesterday's competitive analysis, and it remembers the three competitors you're tracking instead of asking again. You give memory a home that survives restarts.
You take one Hermes agent that's already dialed in and clone it—SOUL.md, AGENTS.md, the skills directory, all of it—into a new folder, then edit from there. One research agent becomes the foundation for a drafting agent and every specialist after it.
The moment your agent stops waiting for you to type and starts running on its own schedule, every loose end you tolerated in testing becomes a standing risk. You tighten the vague boundaries before one fires at 3 a.m. with nobody watching.
Ten instructions, four files. You take each instruction and write its home beside it—AGENTS.md, SOUL.md, MEMORY.md, or state.db—and the four-file table gives you the reasoning behind every best answer. It's the muscle you reuse for everything that follows.
A prompt is a sentence you say once: it works, then it's gone. Three failures follow. It doesn't persist—tell the agent "always cite your sources" today and tomorrow's session has never heard the rule. And it doesn't scale.
When you send a task, the agent doesn't treat these files as separate side notes. It builds one instruction stack from durable inputs: identity from SOUL.md, project context from AGENTS.md, memory snapshots from MEMORY.md, and session data from state.db. That's the part that makes the system click.
Say you want an agent that drafts investor updates. Watch four instructions each land in a different file: "This company is a B2B logistics startup; this quarter's priority is reducing churn" goes to AGENTS.md as project knowledge, while "Write in plain, confident English—never use hype words" goes to SOUL.md.
You score each rule in your draft, and anything scoring low gets rewritten or deleted—an unfollowable rule just adds noise the real rules have to compete with. Start with: is it specific enough to fail a test, or is it just a vibe?
EPUB, PDF, and HTML are included so the book can work on an e-reader, as a designed copy, or as a searchable desk reference.
For e-readers and reading apps.
The designed edition with diagrams and layouts intact.
Searchable, copy-pasteable, and practical as a reference.
Yes. You get the complete edition, including the chapter sequence and internal materials described on this page.
EPUB, PDF, and HTML are included so you can read on an e-reader, keep a designed copy, or use the searchable browser version.
Because this is an instant digital download, broad change-of-mind refunds are not offered after the files have been accessed. Refund requests are reviewed within 7 days for duplicate purchases, accidental purchases before access, access failures we cannot fix, wrong files, corrupted files, or pages that materially misdescribe the book.