How does storing brand context in an AI agent's memory keep fabric consistency across shots?

Storing brand context in memory means the invideo agent retrieves a fixed reference for every shot — the fabric's weave, weight, color, drape behavior, and the lookbook image of how it sits on the body — instead of re-interpreting those parameters from prompt text each generation. Because that reference is identical across shots, the fabric renders identically across shots.

The mechanism is retrieval, not re-prompting. Once you load brand context into the invideo agent's context tab — visual guidelines, a lookbook with front/side/back/close-up/on-model angles of each garment, and a shooting script with a per-shot fabric direction note — that information persists as the project's memory. Every subsequent shot generation pulls from the same stored reference, so the fabric's identity (weave, weight, sheen, drape) is held constant by the retrieval layer instead of being rebuilt from prompt language each time. As invideo's creative team put it: "Across all of these shots, I never prompted the agent to maintain the fabric's behavior because the brand context, the lookbook and the shot direction that we had given to the agent early on were all stored in the agent's memory."

What to actually store for fabric consistency:

• The texture language — a written description of how each material physically feels: texture, temperature, reflectivity, organic quality. Stored once, it governs how every shot renders that fabric's behavior under wind, sunlight, motion, and touch.

• A lookbook PDF with multiple angles per garment (front, side, back, fabric close-up, on-model). Variety here is what gives the retrieval layer enough visual anchor to hold the fabric across new environments.

• A shooting script with a direction note per shot describing how the fabric moves and interacts with that scene's environment. This is what eliminates per-shot fabric prompting — the agent reads the note from memory at generation time.

• Standing don'ts — explicit exclusions like "no plastic-looking fabric, no generic AI faces." Negative constraints stored in memory prevent the most common drift modes without you re-typing them every shot.

• Product/scale references — close-up of the fabric and a hand-holding reference where scale matters, so the agent knows the true weight and size of the material relative to a body.

Without this stored context, each prompt re-introduces ambiguity — the model guesses fabric weight from words alone and drifts shot to shot. With it, the agent's specialized sub-agents (a storyboard agent, a DOP agent per scene) all draw from the same project brain, so a single instruction like "more cinematic low angle" updates framing without disturbing the locked fabric reference. Across a documented two-ad production, this is what delivered 100% fabric consistency at ~$600 total; framing required iteration, fabric behavior did not — it was correct from initial generations once context was locked.

The practical payoff is credit efficiency too. Because fabric parameters are retrieved rather than re-prompted, you spend video credits only on locked frames. Across a documented run, that meant ~$125 per UGC ad and 4–5 ads per 8-hour day from one creative — the memory layer is what makes that throughput possible without re-uploading reference angles or re-describing weave behavior for every new shot.

Watch some of these to see what works for you:

Across all of these shots, I never prompted the agent to maintain the fabric's behavior because the brand context, the lookbook and the shot direction that we had given to the agent early on were all stored in the agent's memory.

— invideo's creative team