Knowledge-vault layer

Agents that forget everything between sessions aren't colleagues, they're calculators. I needed durable memory: decisions, analyses, and runbooks that outlive any single chat, and that every agent checks before it does substantive work.

How it works

A local Obsidian vault, synced across my machines, is wired to Claude through a scoped MCP server. The models read and write durable business knowledge directly. Outputs route by type: tax memos, property notes, decisions, and reusable reference each land in their own subfolder. Binary source documents stay in a separate archive and get referenced by link, not embedded, so the vault stays a clean markdown thinking layer.

• Sensitive folders are read-only to the models, and deletes are gated.
• One source of truth. Agents read prior decisions before acting, then write synthesized outputs back.
• The vault is the memory. Live chat is where the thinking happens, and durable conclusions get filed after.

It's live, and it's the layer under everything else on this map. The ongoing work is discipline more than code: keeping outputs routed to the right folder so the knowledge actually compounds.

This is what turns a pile of one-off automations into a system with a memory. One owner, plus agents that remember what we already decided. Context is the company.