Claude + Hermes Setup: Persistent Memory and Agent OS
Combining Claude with Hermes Agent gives Claude persistent memory across sessions via SQLite, an auto-improving skill system, and scheduled task automation. Hermes can spawn Claude Code as a sub-agent for file and code operations, with Jarvis voice mode allowing hands-free task execution via phone. An Obsidian-based "Agent OS" dashboard ties everything together in dedicated agent rooms with shared memory.
"Claude + Hermes Setup is WILD" by Julian Goldie SEO — Watch on YouTube →
Key Takeaways
- Hermes stores all conversations and skills in a persistent SQLite database that survives across sessions — Claude on its own forgets everything when you close the tab
- Hermes can spawn Claude Code as a sub-agent for heavy implementation tasks, then log results back into the persistent memory system
- Jarvis voice mode uses MiniMax M3 for voice commands — can schedule cron jobs, write files, and execute tasks hands-free over a phone call
- Obsidian Agent OS: separate rooms for Claude and Hermes, shared memory vault, Kanban board, goals page, and content studio in one dashboard
- Learning curve exists: memory configuration and Claude Code permissions integration require careful setup — not plug-and-play
What you can actually set up from this
Extracted from the video's own transcript — the specifics the original summary left out.
Reproducible steps
- Let Hermes delegate heavy coding to Claude Code
Hermes handles memory, scheduling and skills, and hands implementation work to Claude Code. Hermes ships a bundled claude-code skill for exactly this ("Delegate coding to Claude Code CLI").
- Import an existing Claude Code setup
hermes import-agent maps a ~/.claude setup into Hermes — CLAUDE.md/AGENTS.md instructions become memory entries and Bash(...) permission rules become the command allowlist (per the Hermes docs).
- Schedule by voice
He voice-commanded a daily cron job that researches AI automation news and writes a dated JSON file each afternoon.
Gotchas
- Voice wake: with the Porcupine backend the default keyword is "jarvis", not "hey hermes" (per the Hermes docs).
- Browser and screen control are opt-in; understand the permissions before letting it run unattended.
- Where people get stuck, per the video: memory/vault structure, and permissions between Hermes and Claude Code.





