jimmy
9.9 KiB
9.9 KiB
Implementation Guide: Obsidian Second Brain for Your AI Stack (Hermes, Leonard, Charlie, Athena, OmniRoute)
This guide turns the concepts from Julian Goldie's Obsidian 1-Hour Course into concrete, actionable steps tailored to your existing setup (local RTX 3060 stack, Hermes agents, Gitea projects, AI Research Oracle, AI Talk Show, OmniRoute gateway, Phoenix Salon client work).
Prerequisites (You Already Have Most)
- Obsidian app installed on your main machine (or install now from obsidian.md — works great on Ubuntu via AppImage or Flatpak).
- The
obsidian-second-braincross-CLI skill installed and available to your agents (this is the key integration layer Goldie describes). - Local LLM access (LM Studio, Ollama, or your quantized Qwen2.5-14B / Bonsai 27B via ComfyUI/LM Studio).
- Claude Desktop or Claude Code access (recommended for heavy vault ops initially — the $20 tier is worth it for big context).
- Your projects already partially documented in Gitea, SOUL.md, agent memories, chat histories.
Step 1: Initialize the Vault (One-time, 15-30 min)
Goal: Seed the vault with everything your agents already "know" about you so new agents start with rich context.
Recommended Agent: Start with Claude Desktop (best at large vault operations per Goldie) or your main Hermes instance.
Prompt to run first (copy-paste into Claude or Hermes):
You are initializing my Obsidian Second Brain vault located at [INSERT FULL PATH TO THIS VAULT FOLDER].
I want you to act as the chief knowledge architect.
1. Explore the current structure (Projects/, Areas/, Resources/, Archives/, Meta/).
2. Create or update high-level overview notes for my active projects and areas based on everything you know about me from our previous conversations and my profile:
- AI Research Oracle (daily AI news aggregator, Drudge/Citizen Free Press style, using Athena + Leonard)
- AI Talk Show with persistent agents Leonard and Charlie (backgrounds, memories, Gitea project)
- OmniRoute AI Gateway (LLM routing, API key management, best-free failover)
- Phoenix Salon + Spa branding project (with owner Ashley in Cameron Park, CA; hair transformation photos, Phoenix logo watermarks, empowering mottos, social content)
- Local AI/ML stack (RTX 3060 12GB primary, access to RTX 5090 32GB remote, Qwen2.5-14B, Bonsai 27B quantized, LM Studio, ComfyUI for LLMs + NSFW image gen)
- Hermes agent ecosystem and custom pipelines (Athena oracle, etc.)
- My consulting work for small businesses (digital audits, social media strategy, GBP optimization, AI automation via Hermes)
- Ubuntu VPS management via PuTTY/SSH
3. For each major project/area, create a structured note with:
- YAML frontmatter (tags, updated date, updated_by: "Claude-Init" or "Hermes", status)
- ## Summary
- ## Current Status & Next Actions
- ## Key Decisions & History
- ## Related Notes (use [[wikilinks]])
- ## Open Questions
4. Also create/update:
- Meta/Dashboard.md as central Map of Content (MOC)
- Resources/Obsidian-Methodology/ notes if missing
- Any quick-capture Inbox notes from recent activity you remember
5. After creating, give me a summary of what you seeded and any gaps you noticed that I should fill manually or in follow-up chats.
Use the PARA method strictly. Prefer atomic notes + heavy linking over giant walls of text. Make everything agent-readable with clear sections and tags.
Do not ask for confirmation on each file — create them, then report back.
Run this once. It gives you the "really powerful starting point" Goldie emphasizes instead of building from zero.
Step 2: Wire Agents to the Vault (Ongoing)
For Hermes / Custom Pipelines (your primary)
- The
obsidian-second-brainskill should already expose tools like:- Search vault by tag / keyword / semantic
- Read specific note
- Create / append / update note with frontmatter
- List projects by status
- Daily maintenance job
- In your agent system prompts or tool-use instructions, add:
You have persistent memory in the Obsidian vault at ~/second-brain (or your path). Before any complex task, retrieve relevant context using the obsidian-second-brain tools. After completing work that creates new knowledge or decisions, write structured updates back to the appropriate Project or Area note. Prefer daily batch updates over per-message writes to control token usage.
For Claude Desktop
- Simply open the vault folder. Use Projects / Artifacts features or @file mentions.
- Or install community plugins like "Obsidian Claude" if available, or use Advanced URI plugin for deep linking.
For Local LLMs (Qwen, Bonsai via LM Studio)
- Use a strong system prompt template that references the vault path and instructs the model to "think in files" or use provided retrieval tools if your inference server supports it.
- For lighter models, pre-load key notes into context via RAG or manual include for important sessions.
For Leonard & Charlie (AI Talk Show)
- Their character "SOUL.md" or memory files should live in or link to this vault.
- Episode planning, character development, audience Q&A, show bible → dedicated Project note or subfolder.
- After each "episode" simulation or recording, the producer agent (or you) runs an update prompt to log what happened into the vault.
Step 3: Set Up the Daily Infinite Context Loop (The Real Magic)
Goldie's key insight: every chat your agents have gets written back into Obsidian.
But smartly — not raw transcripts (token explosion). Use daily reconsideration/summary.
Recommended Implementation (matches your resource-discipline posture)
Option A: Scheduled Agent (Best — uses your existing cron/Athena patterns)
- Create a daily cron job or scheduled Hermes task (e.g. 6am or after your main work block).
- Prompt:
Run daily Obsidian Second Brain maintenance for [date]. 1. Review yesterday's agent activity logs, conversation summaries, and any new files created in Projects/ or Areas/. 2. For each active project, produce a concise "Yesterday's Progress" bullet list and append or update the project's note. 3. Update frontmatter `updated` date and `updated_by: "Daily-Maintenance-Hermes"`. 4. Create a new daily note in Meta/Daily-Logs/ if it doesn't exist (or append to it). 5. Identify any new connections or tags that should be added. 6. Prune or archive anything that has been completed >30 days ago. 7. Output a short "Vault Health Report": number of notes touched, biggest updates, any recommended follow-ups for me. Keep summaries tight. Focus on decisions, outcomes, new knowledge, and next actions. Use wikilinks liberally.
Option B: End-of-Chat Explicit Write-Back (lighter)
- Add to your standard agent closing instruction: "At the end of this task, if new durable knowledge was created, call the obsidian-second-brain update tool with a structured summary for the relevant project note(s)."
Token Tip from Goldie: Daily summary beats per-message. Your RTX 3060 will thank you. The skill already supports this pattern.
Step 4: Capture from Other Agents (ChatGPT, Claude Web, etc.)
Goldie shows exporting memory from ChatGPT/Claude and importing.
Prompt for export: "Export everything you know about me and my projects into a structured Markdown format suitable for Obsidian. Organize by project/area using PARA. Include YAML frontmatter. Make it comprehensive but clean."
Then paste the output into a new note in Inbox/ or have Hermes/Claude import and file it properly.
Step 5: Backup, Sync & Future-Proofing
- Local first: Your vault lives on disk you control.
- Backup cadence: Full rsync or zipped export every 30 days to external drive + private cloud (or GitHub private repo — but be careful with .obsidian/ folder; usually gitignore it and any large attachments).
- Multi-device: Obsidian Sync (paid) or Syncthing / rclone to your Ubuntu VPS or other machines. Or just git for notes.
- New agent arrives? Point it at this vault folder + give it the skill/tools. It wakes up knowing your entire history, preferences, active projects, and how you like things organized. This is the "slot the vault straight into it" superpower.
Common Pitfalls to Avoid
- Don't manually curate everything — delegate to agents (Goldie's explicit advice).
- Don't let the graph view seduce you into over-linking early. Start with good folder structure + tags + frontmatter. Links emerge naturally.
- Avoid giant monolithic notes. Prefer atomic + MOCs (Dashboard.md is your main one).
- Zettelkasten is elegant but high-friction for fast AI work. PARA + tags wins for your speed.
Your Specific Project Notes to Prioritize Seeding
- Projects/AI-Research-Oracle/...
- Projects/AI-Talk-Show-Leonard-Charlie/...
- Projects/OmniRoute-AI-Gateway/...
- Projects/Phoenix-Salon-Spa-Branding/...
- Areas/AI-Agent-Development/ (covers Hermes, local models, skill development)
- Resources/Local-LLM-Setup/ (RTX 3060 specifics, quantization notes, ComfyUI workflows)
- Resources/Prompt-Engineering/ (your evolving library, especially for agent loops and vault interaction)
Success Metrics (You'll Know It's Working)
- You rarely re-explain context to agents.
- New agents (or new chat sessions) immediately reference past decisions accurately.
- You can search "SEO" or "Phoenix Salon" or "Leonard character background" and instantly see the full picture + recent activity.
- Your Obsidian Graph shows bright clusters around active projects.
- Daily maintenance runs with minimal intervention and compounds knowledge.
This is the unlock Goldie describes: a team of agents that all share the same evolving understanding of you and your work — coordinated by you, not a vendor.
Run the initialization prompt today. Your future self (and your agents) will be dramatically more capable.
If you hit any friction with the obsidian-second-brain skill or want me to generate more specific prompts / note templates, just say the word.