One living memory the whole team builds and draws from. You drop sources and ask questions; agents do the bookkeeping — summarizing, linking, filing, pruning. Knowledge is compiled once and kept current, not re-derived per query.
Zero logistics. Nobody needs a query language, a database, or the filing rules by heart. The conventions live in one schema file (CLAUDE.md) and the agent applies them every time — you read pages, drop notes, and ask in plain language.
You work in one Obsidian vault: a private root wherever you keep your notes — OneDrive, iCloud, plain disk, your call — with the shared libraries mounted inside it. One window, one graph, one search — and nothing leaves your space until you move it. Moving it is the share. Setup is one guided run of /setup-vault: it picks a safe shape and proves the mount with a live sync round-trip.
A note's audience is the container it sits in — never a setting, never frontmatter. Agents synthesize everything they can read, so the only boundary an agent can't leak across is one it can't see. "Who can see this?" = "where does it live?"
Promoting a note is the same deliberate, one-directional act as ingesting any source. Drop it in the container's _inbox/, the agent compiles it — type + compiled, cross-links, you on record as contributor. And shared pages never link into anyone's private space.
Every interaction runs one of four disciplined operations. Each one ends the same way: cross-links updated, index.md refreshed, a line appended to _system/log.md. Ingest and lint are steward-run (plus a scheduled weekly lint); query is open to everyone — and only one writer agent at a time touches the shared library.
Every wiki page declares a type in frontmatter — that's what makes the vault queryable (see Wiki Catalog.base). The typed core is required from day one; the emergent edge appears as good answers get filed back.
The agent keeps these consistent, cross-linked, and current. Don't hand-edit — propose changes through the agent or the page's steward.
Meeting notes, scratch, drafts. Yours entirely — agents leave them alone unless you ask for an ingest.
Ingesting a manual note is a visible state transition — it gains type + compiled and becomes maintained. No copies, no second location.
Raw capture stays in the dev repo; ratification happens at the retro; the vault receives what survived. Knowledge that needs teeth also lands in the repo as a hook or CI check — and the lesson page tracks it.
A domain insight, a client gotcha, a pattern worth remembering — lives only in the vault as a lesson page: what we learned, why, how to apply.
Land in both places: the vault page is the "why", the repo artifact (hook, CI rule, skill) is the "teeth". linked_artifacts pins the repo SHA — when the artifact later changes, the lesson shows drift.