you need to use Hermes RIGHT NOW!! (goodbye OpenClaw!!)

Summary of the video review (Hermes AI agent)

What’s being reviewed

The creator switches from OpenClaw to Hermes—a lightweight, open-source-ish AI agent “harness” similar in spirit to OpenClaw. Hermes is positioned as more stable, less clunky, and more self-improving, with better memory and a stronger skill system.

---

Key features highlighted

1. “Vibe,” mission, and team/philosophy

   • Hermes is described as coming from Nous Research, emphasizing a humanistic, censorship-free, democratic AI mission.
   • The cofounder story frames the origin as hackers on Discord building open-source AI, with a strong focus on product “vibe” and overall experience.

2. Easy installation + flexible hosting

   • Installation is presented as one command (Python-based and lightweight).
   • Can run locally or on a cloud VPS (with Hostinger mentioned as a sponsor).
   • Supports running side-by-side with OpenClaw, including an OpenClaw migration path.

3. Agent setup is streamlined

   • You choose an inference model/“brain” (local or hosted options).
   • Examples mentioned:
     • OpenAI (ChatGPT subscription) via OpenAI account login
     • Grok partnership/subscription option
     • Hosted/local options like OpenRouter and Qwen (local mention)

4. Messaging integration (example: Telegram)

   • Telegram bot setup via BotFather and token configuration.
   • Includes a user gating option so the agent can be configured to only talk to the user (security measure).

5. Memory system that “grows with you”

   • Hermes uses two files:
     • USER.md (about you)
     • MEMORY.md (about the environment/technical context)
   • Differences vs OpenClaw:
     • Hard size limits:
       • USER.md: 1,375 characters
       • MEMORY.md: 2,200 characters
     • Memory updates are actively nudged during the session (described as roughly every ~10 turns), not only at session boundaries/compaction.
     • Memory is curated/distilled automatically to reduce bloat over time.

6. Self-improvement loop + skill system

   • Hermes builds skills based on what it learns from interactions.
   • Example: an agent named Ron Weasley is tasked with setting up Twingate (VPN) from the cloud to enable access to the creator’s studio network.
   • The agent then created a new skill (described in-video as a “Self improvement review skill” / “Twingate client operations created”).
   • Skill philosophy comparison:
     • OpenClaw: skills are found/selected from a marketplace
     • Hermes: skills become repeatable workflows crystallized from interaction

7. Skill curation (“Curator” agent)

   • A background Curator periodically reviews skills:
     • improves them
     • removes irrelevant/stale skills
     • moves skills through active / stale / archive states
   • Goal: prevent skill clutter as self-improvement runs.

8. Built-in skill examples

   • Home Assistant skill:
     • switch lamps on/off, set color, and control automated blinds (creator demonstrates multiple actions)
   • GitHub PR review skill:
     • included in the library; pitched as high-quality based on prior internal PR review experience
   • UniFi management example:
     • provided credentials lead the agent to create a “UniFi network operation” skill via self-improvement
   • Delegation/sub-agents:
     • Hermes can delegate and operate as a functional agent harness (plus the creator notes you can interrupt/steer/queue messages)

9. Newer product additions mentioned

   • Hermes dashboard (skills/plugins/models/profiles, achievements)
   • Kanban (task management)
     • can create and track an artifact (e.g., an HTML “Pokemon card” page)
     • supports human input when blocked
   • Computer use (preview) to control a computer (described with a “shipping/preview” vibe)

---

Pros mentioned (unique positive points)

• Memory grows better and stays focused
  • Hard limits + active in-session updates help prevent clunky bloating.
• Creates its own skills
  • Framed as learning the user’s workflow rather than relying mainly on a marketplace.
• Stable / “doesn’t break”
  • The creator reports OpenClaw became frustrating over time (updates break it, responsiveness issues), while Hermes “hasn’t had one issue.”
• Better fit for non-technical trust
  • The creator claims Hermes was the only agent they felt comfortable giving to their wife.
• Great user experience in terminal and messaging
  • Terminal experience is described as “fun,” not just usable.
• Product philosophy favors simplicity and quality
  • Hermes is framed as intentionally offering fewer/curated options than OpenClaw.

---

Cons / risks mentioned

• AI learning can be scary
  • Concern about the model self-correcting/deciding what to learn.
  • Cofounder response: models are smart and should be allowed to figure out what you want.
• Potential limitations vs “support everything”
  • Hermes provides fewer messaging options than OpenClaw (Telegram setup described as having “less options”).
• Safety comparison to OpenClaw skills marketplaces
  • The creator references prior concerns about OpenClaw community downloads with vulnerabilities/CVEs.
  • No explicit proof of Hermes being more secure—more a “Hermes hasn’t had agent-related hits yet” claim.

---

Comparisons made with OpenClaw

• Clunkiness and degradation
  • OpenClaw felt powerful initially but later became unstable and required troubleshooting.
  • Hermes feels more like a product; OpenClaw feels like a project.
• Memory/bloat behavior
  • Both use user/memory files, but Hermes:
    • enforces stricter caps
    • updates memory more proactively during sessions
    • curates automatically to prevent bloat over time
• Skill approach
  • OpenClaw: marketplace-driven skills
  • Hermes: interaction-driven workflow crystallization, plus a Curator to prune/maintain skills
• Migration/side-by-side
  • Hermes supports an OpenClaw migration path and can run both simultaneously for testing.

---

User experience & setup flow (how it feels to use)

The video shows:

• Installation via cloud VPS
• Agent setup via:
  1. pick a brain/model (e.g., ChatGPT subscription)
  2. choose the terminal backend
  3. configure Telegram bot
  4. start the agent and talk via terminal/Telegram

The creator also describes the installation as “therapy,” followed by quick activation of the agent.

A daily-use scenario is mentioned: the wife’s Hermes agent helps with homeschool/diet planning and home management.

---

Ratings / numerical scores

• No overall star rating or numerical score provided.
• Memory limits are numeric:
  • USER.md: 1,375 characters
  • MEMORY.md: 2,200 characters
• Periodicity claim:
  • memory nudges described as about every 10 turns (approximate wording in subtitles)

---

Unique points mentioned about Hermes (consolidated list)

• Mission/vibe and Nous Research team story (humanistic, censorship-free, democratic AI)
• Built for open-source AI culture; origins as Discord hackers
• Install: lightweight, Python-based, “one command”; runs locally or cloud
• Host integration example with Hostinger
• OpenClaw migration path and ability to run side-by-side
• Terminal UX described as “fun”
• Brain choices: local (LM Studio/Qwen), OpenRouter, OpenAI Codex via ChatGPT subscription
• Grok subscription option
• Telegram setup via BotFather; includes a user-only access option
• Memory system:
  • USER.md + MEMORY.md
  • hard size caps (1,375 / 2,200 chars)
  • curation/distillation
  • active in-session updates (~every 10 turns)
  • memory loaded into the system prompt for new sessions
• “Honcho add-on” mentioned as enabling “peer card”-style personality/context learning (claimed to fit Hermes better than OpenClaw)
• Self-improvement loop:
  • creates new skills from interaction (Twingate example)
  • delegation and sub-agent behavior
• Curator agent:
  • periodically reviews skills
  • improves/removes stale skills
  • manages active/stale/archive lifecycle
• Built-in skills showcased:
  • GitHub PR review skill
  • Home Assistant skill (lights/blinds demo)
  • UniFi network operation skill
• Dashboard features:
  • skills/plugins/models/profiles/aux models
  • achievements
• Kanban task feature (progress + human input when blocked)
• Computer use feature (preview)
• Stability claim: Hermes “doesn’t break,” OpenClaw degrades/breaks after updates
• Trust: creator comfortable giving Hermes to spouse/family

---

Speaker-specific views

• Main creator (NetworkChuck)

  • Explains the rationale for switching
  • Demos installation, setup, memory behavior, and skill growth
  • Compares stability/clunkiness vs OpenClaw
  • Shares personal trust/family use

• Cofounder Jeffrey Quesnelle (Nous Research)

  • Explains company philosophy
  • Discusses the model/learning philosophy (“get out of the way with the models”)
  • Mentions Hermes’ internal origins prior to OpenClaw
  • Addresses safety and the idea of letting models figure things out
  • Adds context on mission and self-improvement/skills

---

Verdict / recommendation (concise)

Based on the video, Hermes is recommended over OpenClaw for users who want a more stable, less clunky agent experience with better long-term memory growth and a self-improving skill system.

The biggest differentiators claimed are active, curated memory updates and skills that crystallize from your interactions, supported by a Curator that helps prevent skill pile-up.