Figma and Lovable learn to stick to what you decided.

Figma: a model for editing, not inventing

On October 29, Figma's VP of Software Engineering Josh Clemm wrote on the company blog about the model Figma is training for edits on the canvas. The Figma agent uses it for precise changes to an existing design; image generation, vectors and new explorations still go to third-party models.

The errors they want to get rid of are the most common defects of AI-made interfaces: icons colliding with text, clipped dropdown arrows, overlapping labels. Plus a subtler one: the post points out that a button can look right and still be built from the wrong component.

The most important line is about training. They say supervised fine-tuning alone copies the training designs too closely and generalizes poorly, so they're adding reinforcement learning with criteria written by expert designers. That's the clearest account I've read from a company of where slop comes from: a model that learned from thousands of similar designs hands you back the average.

The numbers Figma shares are for its agent since it left beta: typical tasks run about 20% faster, user ratings are 33% higher, and people undo its work 60% less often.

Lovable: project memory

On October 30, Lovable's changelog announced that it now remembers the preferences, decisions and facts you state while you build. There are two kinds: project memory, shared by everyone who can edit the project, and personal memory, which follows you across projects.

The docs say project memory can hold design decisions such as colors, typography and tone of voice, along with approaches you've already ruled out. It fills up on its own as you chat, or on request with "Remember that..." or "Forget...", and you can review it under More, Settings, Memory.

What this means for you

Both updates point the same way: AI does better work inside a frame than without one. A model that edits while respecting your components, and a builder that remembers your colors, produce less slop than a tool that starts from zero on every prompt. But neither one decides for you. Figma still sends new explorations to other models, and Lovable can only remember what somebody actually said.

If you never said anything, memory will store the default just as faithfully as it would store your choice. That's the point of who decides the design: whatever you leave undefined, the model defines, and now it also gets written down.

• If you use Lovable, open the project memory today and see what it saved by itself. If something in there isn't a decision you made, tell it to forget it.
• State your decisions outright: "Remember that the typeface is X, the main color is this hex code, corners use this radius, and I don't want icon cards." There's a full brief to copy in the prompt guide.
• If you use Figma, the agent edits better on a file with well-built components. The tidier your library, the less room the average has.

And before you publish, run the result through the Slop Score: memory helps you keep a decision, not tell whether it was a good one.