Knowledge & import
What one agent learns, the next one should start with. The context store is that shared memory — and the importer pulls in the instruction files you already have.
The context store
Context items are short markdown notes — decisions, facts, task state — scoped to a repo or to the whole workspace, optionally visible to one profile only (Work knowledge never reaches Personal agents). Items link to each other with [[wiki links]], keep full edit history, and can be published from anywhere: the context pane, a selection in any terminal, the approval inbox's “publish output”, an automation rule, or aioat publish in a shell.
Every new agent starts with a digest of the relevant items in its opening prompt — the workspace's cross-repo scope plus the repo's own. Flagged (⚠ injection-suspect) items are included but explicitly labelled untrusted.
The context graph
Knowledge → Context graph visualizes it: items, repo instruction files, live agents, and profiles as nodes; links, authorship, and scopes as edges. Click a node to read, edit, or trace who wrote what — and to spot stale or contradictory guidance before it misleads the next agent.
The importer — bring your existing setup
Already have instruction files from other tools — or a whole .claude/ folder full of subagents? Knowledge → Importer scans a repo (or any folder) for:
| Provider | Files |
|---|---|
| Claude Code | CLAUDE.md, .claude/ — skills, commands, subagents (agents/*.md), settings (secret-looking values masked) |
| Codex | AGENTS.md, .codex/ |
| Gemini | GEMINI.md, .gemini/ |
| Cursor / Windsurf / Cline / Copilot | .cursorrules, .cursor/rules/, .windsurfrules, .clinerules, .github/copilot-instructions.md |
Tick what you want and Import — files land in the context store tagged by origin and provider, flow into agents' digests, and appear in the graph. The injection scanner runs on everything first; warnings show before you import.
- Drift: if a file changes on disk after import, its row shows ⟳ drifted — re-import replaces the stored item (history keeps the old version).
- Overlap: instruction files from several providers in one scope get a warning — agents receive all of them, so review for contradictions.
~/.claude/agents/freja.md don't need recreating. Copy them into the repo (or your local .claude\agents\), and every Claude session finds them automatically — import them too and AIOAT's other agents get to know your team exists.Export the canon
Maintain guidance once, ship it in every tool's native format: Export canon as… writes the scope's instruction items back out as CLAUDE.md, AGENTS.md, GEMINI.md, .cursorrules, and friends. Exports carry an AIOAT marker — the scanner won't offer them back as imports, and exporting over a file AIOAT didn't write asks before overwriting.
Side chat
A lightweight Claude session for questions that don't deserve a full agent — seeded with whatever context you send it (select text in any terminal → Send to side chat, or aioat send from a shell). Promote a side chat into a real agent when it turns into work.