本文由 AI 分析生成
建立時間： 2026-05-30 來源： https://x.com/sourfraser/status/2035454870204100810
Summary
A practical guide to building a persistent AI assistant using Obsidian as a knowledge base, Claude Cowork as the intelligence layer, and MCP connectors to bridge tools like Slack, Google Calendar, and Google Drive. The key insight is that persistent context (a growing vault of business knowledge) transforms Claude from a “temp worker with amnesia” into an employee who compounds knowledge over time. The system automatically processes call transcripts, extracts decisions and action items, and files them into the correct Obsidian notes.
將 Obsidian 作為知識庫、Claude Cowork 作為智能層的持久 AI 助手建置指南。核心洞察是持續增長的情境（業務知識庫）將 Claude 從「失憶的臨時工」轉變為每天變得更懂你的員工。系統自動處理通話記錄並提取決策和行動項目。
Key Points
- Memory file is the “onboarding doc” — contains who you are, business structure, processes, tools, communication style, and goals
- Fathom + Zapier → Google Drive pipeline: every call becomes a searchable transcript automatically, zero manual steps
- Claude Cowork + Obsidian MCP: Claude reads/writes directly to the vault; no copy-pasting required
- System compounds over time: Week 1 knows basics; Week 8 catches things you missed, connects dots across domains
- Store the Obsidian vault in Google Drive for multi-machine sync (native Obsidian support)
- One critical instruction: “Before answering any question, always search the Obsidian vault for relevant notes”
Insights
The most underrated part of this setup is that the Memory file serves as a forcing function to document institutional knowledge that otherwise lives only in founders’ heads. The compound effect described (vault gets smarter every day) is real because the bottleneck for LLM usefulness is context, not model capability. This is a concrete implementation of the “LLM Wiki” pattern (Karpathy’s Andrej’s pattern) where the AI maintains a persistent, compounding knowledge base rather than RAG-style re-derivation on every query.
Connections
Raw Excerpt
The vault grows every day. And because Claude reads it at the start of every session, it knows more every time you talk to it.