All nine Hermes configuration areas in depth: setup location, three-tier model routing, messaging platforms, eight memory providers, skill system, file structure for multi-project setups, agent profiles, delegate task vs profiles decision, and cron job plus Kanban workflows.

Watch on YouTube

Setup Location

Options: main laptop, Mac Mini server, or VPS. VPS recommended: runs 24/7, cheap, can be destroyed and recreated, can be hardened. Hetzner is preferred over Hostinger (Hetzner is month-to-month, Hostinger requires 2-year commitment). A basic Hetzner VPS under $10/month is sufficient.

Three-Tier Model Strategy

Hermes uses three model tiers:

Tier	Role	When
Primary	Conversation and reasoning	Your best model - what you interact with
Auxiliary	Background compression and summarization	Does NOT need to be frontier. Tell Hermes: "Use GPT-mini for auxiliary tasks."
Fallback	Provider exhaustion backup	Switches if primary exhausts credits or rate limits

Fallback rotation strategies: fill-first (exhaust first account then switch), round-robin (alternate), least-used (always use account with most remaining capacity). Configure via hermes auth.

Memory System

Eight memory providers exist. Start with static files.

Static files (loaded every session):

• soul.md: agent persona and communication style. Keep under 60-100 lines. Every line costs tokens on every session.
• user.md: who you are, your business, your preferences. Keep under 60-100 lines.
• memory.md: key facts the agent should always have. Keep under 60-100 lines.

Dynamic memory:

• Daily memory: log file the agent maintains automatically
• Holographic (recommended starting point): local SQLite, free, self-hosted. Activate: hermes memory setup holographic, select yes to extract sessions, use defaults, start new session.
• Mem0: cloud provider, good reputation
• Honcho: mixed results reported

Each profile has its own memory and soul files. When creating a new profile, configure its static files separately.

Skills System

Skills are reusable workflows in SKILL.md format. Key properties:

• Progressive disclosure: not loaded unless relevant to current task
• Self-improving: after several uses, Hermes asks itself whether to update the skill. Correct it once; it updates automatically.
• Cross-harness portable: take a skill from Claude Code, drop it in Hermes, it works.
• No effective limit: 500 skills work fine due to progressive disclosure

Multi-Project File Structure

projects/
  agent.md           - central overview for all projects
  client-1/
    agents.md        - project-specific instructions
    execution/       - deterministic scripts for this client
  client-2/
    agents.md
  investment/
    agents.md

Start Hermes from projects/ as CWD. It has overview of all projects but stays within active project's scope. Strengthen in soul.md: "Do not step outside the currently active project."

Do not install Hermes multiple times for multiple clients. Use file structure and CWD binding.

Agent Profiles

Each profile has its own: memory, gateway, auth providers, model, soul.md. A profile is effectively a different agent instance.

Create profiles by telling your main agent: "Create profiles for coder, researcher, and writer. Copy my auth providers to them, use GPT-5.5 for all of them, populate basic soul.md files."

Start with 1-3 profiles. Have the main agent delegate to specialist profiles.

Delegate Task vs Profiles

Pattern	Type	Memory	Use For
Delegate task	Ephemeral	None	Parallel one-off tasks, fan-out to 5 things simultaneously
Profile-based	Persistent	Own memory	Recurring workflows, role-specific context, different models per role

Delegate task is faster in the moment. Profiles are more reliable for recurring tasks.

Workflows: Cron Jobs and Kanban

Cron jobs: Natural language scheduling. "Deliver me a daily stock summary at 7 AM EST. Schedule this as a cron job."

Kanban board: hermes kanban init then hermes dashboard. Visual task board with models used, logs, cron jobs, memory status, and the Kanban board. Use for complex multi-step tasks with dependencies.

Self-Improving Research Agent Example

Tell Hermes: "Create a research agent that goes daily and researches news and videos about stocks that would interest me. Every day bring me five video suggestions. I will rate them. Use Kanban to organize this. Set a 7 AM EST cron job."

Over 7-10 days of rating results, it refines what it delivers. This is the feedback loop that makes Hermes genuinely self-improving for your specific use case.

Firecrawl integration recommended for any Hermes instance: bypasses Cloudflare, handles most scraping needs, free 1,000 credits to start.