--- title: Taste: teaching a system what a brand likes canonical: https://withicarus.com/papers/taste site: Icarus description: Taste stores what a brand keeps choosing as words, never as pixels. The entry format, the four write paths, why recall lists before it ranks, and why the image an entry was learned from is removed from the render request in code. --- # Taste: teaching a system what a brand likes Taste stores what a brand keeps choosing as words, never as pixels. The entry format, the four write paths, why recall lists before it ranks, and why the image an entry was learned from is removed from the render request in code. Published 25 September 2026. Written by Tanmay (https://tanmay.xyz). ## Abstract Taste is a per-brand store of what a customer keeps choosing, held as prose rather than as pixels. Each entry is one image read back into a paragraph, keyed to the shot it came from, tagged with a shoot type and linked to a re-hosted copy of the original. Four write paths keep it current and one of them removes contradicted entries. Recall lists a container for coverage and then re-orders it by vector search, because using the search for both silently dropped whatever it ranked low. The image an entry was learned from is shown to the director and removed from the render request in code, so the engine rebuilds a look from its description and never copies it. This note covers the entry format, the four writes, the recall path, the bounds in table 1, and four bugs that each moved a design decision. ## Contents 1. [Competent and anonymous](https://withicarus.com/papers/taste#problem) 2. [What one entry is](https://withicarus.com/papers/taste#entry) 3. [Why the read is prose, not a schema](https://withicarus.com/papers/taste#prose) 4. [Four ways it gets taught](https://withicarus.com/papers/taste#writes) 5. [Where it lives](https://withicarus.com/papers/taste#tiers) 6. [Recall: list for coverage, rank for order](https://withicarus.com/papers/taste#recall) 7. [A lane narrows, it does not hint](https://withicarus.com/papers/taste#lanes) 8. [Recreate, never copy](https://withicarus.com/papers/taste#constraint) 9. [Taste and Wings together](https://withicarus.com/papers/taste#wings) 10. [What the settings panel had to learn](https://withicarus.com/papers/taste#ui) 11. [What this does not fix](https://withicarus.com/papers/taste#limits) 12. [Components](https://withicarus.com/papers/taste#built-on) ## Questions ### Does Icarus copy the reference images I save? No. A saved look is read back into words, and only those words reach the image engine. The image itself is removed from the request by the code that assembles it, on every path that spends credits, even though the file is hosted on our own storage and would otherwise pass every check. The engine recreates the look from the description; it never receives the pixels. ### How does Icarus learn my brand's style? Four ways. You pick aesthetics during onboarding, you like or save shots as you work, you can bulk upload up to thirty reference images at once or scan an existing site, and you can dislike a look to have it forgotten. Every one of those writes an entry describing what the image actually was, not the brief that produced it. ### Where is taste stored, and who can see it? In three separate scopes: your own, your workspace's shared default, and a house baseline underneath both. Recall uses the first one with anything on file. Your personal library is yours to clear in one action; the shared workspace one cannot be cleared by a single member, only pruned look by look. ### What happens if I ask for a shoot type I have no saved looks for? Icarus says so rather than quietly using your other taste. It shows you a few saved looks and asks which belong to that shoot type, and it does not spend credits on a render until you answer. Silently substituting nearby taste is what produces a shot that is confidently the wrong thing. ### Do more reference images make the result better? For taste, more saved looks help, because recall ranks them and uses the closest ones. For pinning a specific product or face it is the opposite: one photograph per subject. Multiple angles pull the subject in different directions and it drifts across a set. ## References - [Supermemory](https://supermemory.ai) -- the memory store and vector search taste recall is built on. - [fal](https://fal.ai) -- runs the image and video reads, and the rest of the model calls. - [Inngest](https://www.inngest.com) -- durable background jobs; the enrichment read runs as one. - [Vercel AI SDK, issue 8180: tool-result image parts are dropped](https://github.com/vercel/ai/issues/8180) -- why taste images are sent as message parts rather than returned from the tool. - [Schema.org TechArticle](https://schema.org/TechArticle) -- the structured description this page publishes about itself. ## Read it The full paper, with figures and tables: https://withicarus.com/papers/taste --- Site: Icarus (https://withicarus.com) Canonical: https://withicarus.com/papers/taste Contact: tanmay@onai.studio