The assistant gets better the more you work with it, through two kinds of durable knowledge: memory — facts and preferences it recalls in later chats — and skills — reusable build patterns it captures from successful work.
Memory
Tell the assistant something worth keeping — “remember that our brand font is Inter”, “I always want 9:16 for social” — and it stores it as a memory. It also saves notable preferences, decisions, and lessons on its own as you work together, and recalls the relevant ones in future chats so you don’t repeat yourself.
Memories come in four kinds — preferences, decisions, lessons, and facts — and each carries an importance level that shapes what gets recalled first. They’re scoped to the current project and stored in your browser.
To manage them, ask directly: “what do you remember about this project?” surfaces stored memories, and you can tell it to forget or correct any of them.
Skills
A skill is a reusable subgraph template — a build pattern like “product hero shot: text → image → video” that the assistant captured after it worked. When a later request matches, it replays the pattern instead of designing the wiring from scratch.
Skills follow your account, not the project or browser: they’re available in every project and on every device. Two ways they come to exist:
- Ask for one after a build you like: “save this setup as a skill for product launches.”
- The assistant captures one itself after a verified, successful build.
The skill library
The book icon in the chat panel’s header opens your Skills library. From there you can:
- Search skills by name, description, or when they apply.
- Share a skill — send a copy to a teammate by email. Skills shared with you show who they came from.
- Delete skills you’ve outgrown.
Name the moment, not the mechanism. A skill saved as “weekly newsletter header” gets matched to future requests far more reliably than one called “text plus image setup”.
