Taste memories
The eight categories Krafa learns from a board, the shape of a memory, and what separates a useful one from a useless one.
A memory is one durable design principle Krafa believes about you, backed by evidence from something you actually pasted. The set of active memories on a board is that board's taste profile, the thing your coding agent reads.
Open the panel with the wand button in the top-right of any board, or ⌘K → View Taste Memories.
Principles, not descriptions
The distinction that makes the whole product work: Krafa records what a reference implies about your preferences, not what the reference is.
Paste a screenshot of Linear's issue list and a system that described artifacts would store "a dark issue list with a sidebar." Useless, because your agent isn't building Linear's issue list. What Krafa stores instead is the transferable part:
Spacing & Density: Dense list rows favor a tight vertical rhythm with generous horizontal gutters
row-height: 32px·padding-inline: 16px·no dividers between rows
That's an instruction. It applies to a settings table, a file browser, a leaderboard, anything with rows, which is the point.
The eight categories
Every memory belongs to exactly one:
| Category | Covers |
|---|---|
| Typography | Families, scale, weight, tracking, line height, hierarchy |
| Color | Palettes, accent usage, contrast, surface treatment, theming |
| Spacing & Density | Rhythm, padding, gutters, how tight or airy things sit |
| Layout | Grids, composition, alignment, page structure |
| Imagery | Photography, illustration, iconography, graphic treatment |
| Motion | Transitions, easing, duration, what animates and what doesn't |
| Voice & Tone | Microcopy, labeling, capitalization, how the product talks |
| Components | Buttons, inputs, cards, and other recurring UI patterns |
The panel groups memories under these headings in a fixed order, and so does the Markdown profile your agent reads. An agent looking for "what typeface should I use" finds it in one place instead of scanning a flat list.
Anatomy of a memory
Prop
Type
Evidence is what keeps a profile honest. A statement alone is an opinion; a statement with accent: #ff5c00 next to it is a specification. It's also what makes a wrong memory easy to spot: if the evidence doesn't look like something you'd ship, the statement is wrong.
Internally each memory also tracks which board nodes produced it, which is how deleting a reference retracts the belief and how exemplar images are chosen.
What a good profile looks like
A useful profile is small and specific. Twenty sharp memories beat two hundred vague ones, and Krafa is deliberately conservative, writing at most three memories per reference rather than mining every paste for everything it can say.
Signs a profile is working:
- Statements read as instructions you could hand a designer
- Evidence contains values, not adjectives
- Nothing in it surprises you
Signs it needs attention:
- Statements that only make sense for one specific screen
- Categories that are empty because you've only pasted one kind of reference
- Anything you'd argue with, which is your cue to edit or delete it
Motion and Voice & Tone stay empty on most boards, because screenshots don't carry either. Paste a screen recording's source URL, a product's onboarding copy, or a snippet of your easing tokens if you want your agent to get those right.