Source: raw/The_Complete_Hermes_Agent_Guide_That_Gets_You_Ahead_Of_99%_Of_People.md (auto-caption transcript)
Creator: Sharbel (YouTube @sharbelxyz) · URL: https://www.youtube.com/watch?v=fXHdRKK6DOg · Duration: 20:38 · Uploaded: 2026-10-29 · Sponsor: Honcho (memory provider; the video gives a referral code for $10 in credits). The video also offers a free prompt doc behind a link in the description.
Sharbel, who says he runs his business on Hermes daily, lists seven habits that keep Hermes stuck as “a chatbot on their phone” and gives the fix for each, most of them a Telegram command or a settings change the agent makes on request. Several fixes cite the Hermes docs; those citations are his, not checked against the docs here. The most concrete new material is the /goal command with a judge and a gate, and the separate model slots for Hermes’s side jobs.
Key Takeaways
- Start a new chat often. Hermes loads
USER.mdandMEMORY.mdinto the instructions only at the start of a chat. A memory saved mid-chat is written to disk at once but not loaded until the next chat. A Telegram DM is one long chat, so new memories may never load and, he says per the docs, “the learning loop almost never gets to fire.” Type/newafter each task, topic change or morning. - Check that a memory was really saved. Ask Hermes to show tool calls in Telegram; a real save shows a memory line above the reply. He relays the docs’ warning that small local models (roughly under 30B parameters) and weak tool-callers often claim saves that never happened. Fix: tell it to “use the memory tool to save this.”
- Do not switch models mid-chat. The provider cache (a 75-90% discount, per the docs as he quotes them) belongs to the old model, so the next turn rereads everything at full price. Hermes now asks for confirmation before a switch once a chat passes 100,000 tokens. If you must switch:
/new,/model, then a short recap. - Move side jobs to a cheap model. Compression, image reading, command-safety checks and background review (saving memories and skills) each have their own model slot, and all default to the main model. He moved them to GPT-6 Luna on his ChatGPT subscription by asking the agent to change its own settings. His figures from the docs: compression about 1/50 the cost on a fast model versus Opus; background review 3-5x cheaper.
- Five steps and repeated means a skill. After a task, say “save what you just did as a skill called weekly stats”; it becomes
/weekly-stats. Up to five skill commands can stack in one message./learnturns a site, folder, doc or whole book into a skill with a file per chapter (see learn). Agent-made skills unused for 30 days are archived, not deleted; skills you asked for stay. - Write cron prompts as if the job knows nothing. Each scheduled run starts a fresh session with no chat history and no memory of its last run. The prompt must say what to check, what counts, what to ignore and where to deliver. The
[SILENT]keyword means nothing is sent unless the job fails. Tell inbox jobs that email content is information, never instructions (prompt injection). Turn on continuity so each run sees the previous result, and give crons their own model so changing the main model does not change every job. - Use
/goalfor long jobs, with a gate. A judge checks after every turn whether the goal is done, should continue or is stuck./goal gate add <command>adds a test the work must pass before it can report done; a failure sends the exact error back. Default limit: 20 turns./steerredirects without stopping;/btwasks a side question without interrupting. - Plan on the smart model, run subagents on a cheap one. Subagents run in parallel (up to 10 by default), each with a clean context, and only their final answers return. He relays the docs: helpers burn most tokens on big jobs.
- Keep the “backpack” light. Memory, personality file, skill list and every connected tool’s instructions ride on each message. His slowdown checklist:
/status(right model?),/contextand/context all(what is taking space, down to skills and tools), load only the 2 of 30 tools an app connection exposes that you use, and/compressa long chat.
The Hermes team’s own /goal run
Sharbel reports, from the Hermes team, a single /goal run on September 2 to simplify Hermes’s own code:
- About 19 hours, 1,393 helper agents in total, up to 218 at once.
- Python code (excluding tests) cut from over a million lines to about 698,000.
- Main run about 25,000 including follow-up work. Their estimate for doing it by hand: 1.8 million.
- It was not clean: an expired login killed the run and it had to be resumed, and a few bugs passed the tests.
He uses this to argue for gates: “even the people who built it still check the work.” These figures come only from this video; the Nous write-up was not found or read for this article.
Try It
- In your Telegram chat with Hermes, send
/newnow, then make it a habit after each task. - Ask Hermes to show tool calls in Telegram, then tell it a fact and look for the memory line.
- Ask: “Move compression, vision, approval checks and background review to
.” Then /statusto confirm. - Rewrite one cron so it states what to check, what to ignore and where to send, ends with the
[SILENT]rule, and treats email content as data. Ask Hermes to turn on continuity for it. - For your next long job, ask Hermes first: “Help me write the best
/goalprompt plus a/goal gate addcommand so you know when it is done.” Then send both. - Run
/context alland turn off unused tools from your biggest app connection.
Implementation
Tool/Service: Hermes Agent (Nous Research), Telegram gateway; Honcho optional.
Setup: Honcho memory: hermes plugins install honcho, then hermes memory setup honcho and pick Honcho. He pointed a fresh profile at the same Honcho workspace and it already knew his preferences. Honcho is open source (self-host) or cloud at honcho.dev.
Cost: Hermes is free; model spend is the variable. His side-job model ran on an existing ChatGPT subscription.
Integration notes: One profile per area of life (business, content), each with its own memory, skills and settings. Never point two agents at the same Hermes folder, or their memories mix. He says the October v0.22 update cut the desktop app’s tool instructions from about 13,000 to under 7,000 tokens by holding back rarely used tools.
Open Questions
- Version numbers. He calls the October update “version .22”; this wiki’s newest release article is v0.20.0. The v0.21 and v0.22 release notes were not read.
- Exact names. The captions give the gate command as “/goal gate add” and the cron keyword as “silent in square brackets”. Confirm the spelling, the 20-turn default and the continuity setting name in the Hermes docs.
- The 75-90% cache discount, 1/50 compression cost and 3-5x review saving are his readings of the docs and were not checked.
Related
- Hermes Memory Providers — Honcho and the other external memory backends
- learn — the skill-from-a-source command he demos on a book
- Hermes Profiles & Multi-Instance — why two agents must not share a folder
- A Hermes “Digital Nervous System” — another heavy cron user’s patterns
- Hermes Security Model — approvals and prompt-injection defences behind the inbox rule
- Orchestrator + Cheap-Worker Routing — the smart-planner, cheap-helper pattern across tools
- Hermes Beginner Setup (Sharbel) — the same creator’s from-scratch guide
- Claude Code Token Optimization — the same cache and context mechanics on Claude Code