Designers have an amnesia problem too.
We are the case study
This project has had four names in under a year. figbud shipped its first commit in July 2025. The modern engine started clean as Ark on 2026-03-23. One day later it was renamed Noche. One day after that, Memoire. Two renames inside 48 hours.
Every one of those eras worked on the same problem with the same design sensibilities. And before the workspace store existed, almost none of that knowledge moved forward on its own. The same design system got re-extracted, the same component conventions got re-explained, the same prompts got retyped from memory. The agent did not remember, and honestly, neither did the repo.
So when I say design agents have an amnesia problem, this is not a hypothetical user persona. We were re-teaching our own tool our own system, repeatedly.
The genius without a hippocampus
A model can read a Figma file once and reason about it brilliantly. The reasoning evaporates the moment the session ends. Next session, next project, it does the same work over again. Costs add up linearly. Nothing compounds.
The bottleneck is not reasoning. The bottleneck is memory in a form humans and agents can both read and trust.
What a Note is
A Note is a folder with a note.json manifest and one or more markdown skill files. It tells the agent:
- What this skill does. (“Pull a design system from a URL and produce DESIGN.md plus tokens.json.”)
- When to invoke it. (
activateOn: { intent: "extract-design-system" }) - How to do it. (Step-by-step instructions, code snippets, validation rules.)
When the agent classifies a task and the activation context matches, the Note is injected into the prompt. The agent does not re-discover the workflow. It runs the proven version.
The growth curve is lumpy, and that is the point
You can trace the Notes pack through git, and the shape of the curve taught me more than the count.
The Notes page went up on 2026-03-26, when the homepage still had ASCII flowers on it. Two days later the pack stood at 27 Notes at v0.3.0 (commit 23ad4df, 2026-03-28). Then it crawled: a single docker-environments Note made it 28 on 2026-04-06, and a pass on 2026-04-08 took the count from 28 to 32.
Then on 2026-05-07 one engine commit added 19 agent and motion Notes at once (commit 20a0d3b9).
That lumpiness is the tell. Notes did not accumulate on a content calendar. They graduated in bursts, when a stretch of real work ended and the workflows that survived it were worth writing down. The 19-in-one-commit batch landed the same week the v0.17 features and Studio shipped, because that stretch of work generated 19 proven moves. A memory library that grows smoothly is probably being filled. One that grows in bursts is probably being earned.
Humans have to trust the memory too
The other early lesson: agent-readable was the easy half.
Two days after the pack reached 27, we rebuilt the Notes page for scanning: a 3-column grid, category filters, and click-to-expand cards that show each Note’s activation context and its freedom level (commits cc5e388 and 77ef7d7, 2026-03-28).
Activation context and freedom level are the two questions a skeptical designer actually asks: when does this thing wake up, and how much rope does it get? If a human cannot audit those in ten seconds, they will not let the Note anywhere near their prompt. Memory that humans cannot inspect does not get adopted, no matter how well agents consume it.
Categories that ship
Mémoire ships four built-in Note categories: craft (mobile design, systems thinking, accessibility), research (competitive analysis, data synthesis, surveys), connect (Notion, Linear, Slack integrations), and generate (React Native, Vue, SwiftUI, Flutter codegen).
Each category is open. Each Note is a markdown file. Each one is documented in the community Notes repo.
”So we let it read.”
That line has been on the homepage since 2026-03-26, the first week the site existed, back in the flower era. The memory pitch predates almost everything else here: the Studio app, the harness, the workbench. We knew amnesia was the problem long before we had most of the cure, because we kept paying for it ourselves.
Where this goes
The pack is small on purpose. The early Notes should be the ones that paid off in real work, graduated into something reusable, and survived a second run. If you have a workflow you keep redoing, write it up. Open a PR.
The durable lesson from our own four-name history: an agent’s intelligence resets every session, so anything you want to compound has to live in a file. The future of agentic design is not “the agent figures it out every time.” It is “the agent has a library of proven moves, graduated by real practitioners, and picks the right one.” Notes are the substrate. Help build the library.