Skills-driven content calendar agent — web + email trigger
Everything here is injected into every future request's prompt (see prompts.py / memory_store.py). standing_rules, verified_dishes, and banned_terms are curated — add/remove them as lessons get confirmed. request_notes auto-fills whenever feedback is submitted; prune anything that turned out to be one-off rather than a general rule.
Durable behavioral instructions applied to every request.
Dishes confirmed real — skips re-verification.
Known-fake terms the agent must never use again.
Auto-logged raw feedback from past requests (audit trail).
Nothing here yet.