A centralised knowledge substrate that holds the durable context behind Initial Infrastructures - decisions, contacts, projects, frameworks, market intel, processes - and that the LLM agents from the other in-house projects (Initial Agents, Crawlswarm) point at to ground themselves before they act.

Source of truth: a structured Obsidian vault, versioned with Git, written entirely in Markdown with a strict YAML frontmatter convention (title, source, date, tags, type). Twelve top-level directories with explicit semantics - inbox, projects, knowledge, people, meetings, daily, weekly, monthly, frameworks, references, business, archive - so every note has one obvious home.

Multi-source ingestion: scripts pull external content into the vault (web articles for now, with the same pattern designed to extend to GitHub, Gmail, YouTube, Discord). Each ingested item lands in the right directory with filled frontmatter, ready for refinement.

Claude Code skills: custom skills organised into three families - assistant (meeting prep, decision logs, project status, contact briefs, research, ask), business (ad copy, client tracker, coaching sessions, content repurpose, funnel briefs, offer builder, swipe file, weekly growth), and structuration (braindumps, daily brief, weekly synthesis, monthly consolidation). Each skill is a structured prompt that turns a free conversation into a properly-classified, properly-formatted vault entry.

Proactive capture: an explicit rule embedded in CLAUDE.md instructs the assistant to watch for durable information during conversations (decisions, facts, contacts, processes) and propose saving them, with deduplication checks and a destination decision tree before any write happens.

Why it matters for the rest of the stack: Initial Agents and the other in-house projects increasingly rely on this vault for grounding. It is the slow, curated memory layer underneath the fast, transactional agent runs - the place where a decision made once stays available the next time an agent needs it.

Status: in active use, growing every week. Will harden into a proper service layer (exposed via an MCP server) as the agent footprint grows.