remember stores a standing rule; it is not a brand kit

remember stores a standing rule. "Always 9:16." "Audience is Gen-Z gamers, keep the tone playful." "Keep captions minimal." It is not a brand kit. It is not a one-off for this request. It is not an upload card for a face. Save a logo description here and the renderer will not get a file. Save "make this one darker" here and every future clip arrives darker until you find it in Saved Context and delete it.

Teach the agent a lasting preference is the named agent job. Say it once; it sticks. Required: content and category. The user can view and edit everything saved in Saved Context settings. Saving does not use generation credits. The tool description is explicit: do not use it for one-off instructions scoped to the current request, and do not use it for brand assets (colors, fonts, logo, tone) — those belong in set_brand_kit.

What the job is

Durable preferences and lasting facts about you or the business that future conversations should know. Category-tagged. Applied going forward, including new threads. That is why visibility matters: a three-week-old memory still shapes today's routing whether anyone is thinking about it.

Use it for decision rules: default duration family, caption density, which chat brain to plan with, words you will not claim, "when I say caption I mean burn-in, not a social caption." Those sentences have no kit field, and they should not have to be retyped.

What it is not

• Brand kit writes. Set my brand kit holds values applied literally. Hex, font, logo URL, tagline, stored tone. The split exists because a hex should not be re-parsed out of a paragraph of memories every generate. Put tone in the kit when it is a brand voice to apply; put "playful because the audience is gamers" in memory when it is a targeting rule. If you are unsure and it has a file or a hex, kit first.
• Reusable characters. Set up reusable characters and products is request_asset_upload. A memory of "our mascot is a blue fox" will generate a new fox. An asset will not.
• This request only. "Make the lighting warmer" on the last still is edit_previous_generation or a follow-up prompt. Remembering it teaches the agent that you always want warmer lighting. That is a different product. Confirm you mean always before this tool runs.
• A generate, a post, a schedule. Memory changes the prior on future jobs. It does not spend a model and it does not ship a file. If you wanted a clip, name a create job.

Credits wasted on this row are almost never the save (the save is free of generation credits). They are the later generates that inherited a sloppy memory, or the later generates that never inherited a hex you thought you had stored.

The test

Will this sentence still be true on the next thread, next week, with no file attached?

Yes: this job. Content plus category. Check Saved Context once so you know it landed.

If it is a hex, logo, or typeface, kit. If it is a face or SKU, asset. If it is only for the clip in front of you, do not remember it — just ask.

FAQ

What happens if I say "remember" about a brand color?

It may land as a memory paragraph instead of a kit field, and then it will not merge like a hex. Ask for set my brand kit. The agent is supposed to offer that when brand details show up as asides; make the offer explicit if it does not.

Can I remember a default video model?

As a preference, yes — "always plan talking heads on this row." It still does not pin a renderer the way a saved workflow does. For a series, put the model in the workflow. Memory is a prior, not a lock.

How do I undo a bad memory?

Saved Context settings. Edit or delete. Do not layer a contradictory memory on top and hope the newer one wins every time.

Does teaching a preference generate an example so I can see it?

No. The next create job will. If you need to verify, run a cheap still with the kit and memory on, not a hero video.