Publishing and performance

    Agent memory

    Also called Long-term memory, Lasting preference.

    Agent memory is a lasting preference or fact the agent stores across conversations, separate from the structured brand-kit fields it applies to a render.

    The remember tool stores standing instructions: always vertical, the audience is Gen-Z gamers, never a stock voiceover. They survive the next conversation. They are not pasted into a render as a hex value or a logo file; they shape how the agent works with you.

    Brand-kit is colours, fonts, logo, tone, default ratio — values the generation actually consumes. A prompt is one job's brief. A seed is a sample from a distribution. Putting a logo into memory, or a 'never do X' rule into the kit, files the fact where nothing will apply it.

    Review and edit the store in Saved Context. One-off instructions for this request do not belong here; that is just the request.

    In practice

    • Save rules about how to work — ratio, audience, what never to do — not files the render has to fetch.
    • Put colours, fonts and the logo in the brand kit, even if you are tempted to 'just remember' them.
    • Audit the store when output starts ignoring a rule; stale or conflicting memories are applied too.

    The mistake to avoid

    Saving a logo or a hex into memory, or a standing 'never do this' rule into the brand kit. Each store only applies what it was built to hold.

    Go deeper

    Long-Term Memory: Preferences the Agent Keeps Between Sessions

    Two persistent stores exist: memory for durable preferences, brand kit for visual assets. People put things in the wrong one constantly. Here is the real line.

    Related terms

    The all-in-one AI studio for creators. 60+ models for video, image, voice, music and lipsync in a single app.