memi

Read-only design engineering for coding agents.

Memi is the read-only design engineering audit and skill layer for coding agents. It gives Codex, Claude Code, Cursor, Grok Build, and MCP clients file-anchored interface evidence before they edit UI.

Memi reads the product you already have, identifies accessibility, hierarchy, state, responsive, motion, and token risks, then reruns the same deterministic check after a fix. Your code remains the source of truth.

Supported today: Node 20, 22, and 24 on macOS, Linux, and Windows. Figma and Studio are optional companions.

Links: npm · memoire.cv · current versions · MCP Registry · Agent Skills

Quickstart

Run one audit in any frontend repository:

npx -y @memi-design/cli@latest diagnose . --json --no-write --fail-on none

You get a score, normalized finding IDs, confidence, provenance, and file:line evidence. No account, API key, Figma file, global install, or daemon is required.

Keep the workflow available to your coding agent:

npx skills add memi-design/memi --skill audit-frontend-design

Then ask:

Audit this frontend before editing it. Prioritize the five fixes that will matter most to users.

Focused skill	Use it when
audit-frontend-design	Find interface risks before changing UI
remember-design-system	Load compact product-system context
enforce-design-ci	Gate pull requests with deterministic evidence
build-swiftui-interface	Build and verify native Apple interfaces

Compatible with the shadcn registry and v0 design systems.

If Memi catches a real interface issue in your project, star the repository and share the finding. That is the most useful signal for deciding what to improve next.

What Memi catches

Signal	Example evidence
Accessibility	Missing labels, reduced-motion fallbacks, focus and contrast risks
Interface craft	Weak hierarchy, spacing drift, brittle responsive behavior
Product states	Missing loading, empty, error, success, and permission states
Design systems	Token drift, raw values, inconsistent component usage
Routes and components	App graph, duplicated patterns, risky change surfaces
Pull requests	New debt only, SARIF annotations, step summary, HTML report

The first command is deliberately read-only. Write-capable scaffolds and Figma operations are separate, explicit workflows.

How it works

1. Inspect - build an evidence graph from source, routes, styles, and local design-system files.
2. Find - report normalized issues with severity, confidence, provenance, and file:line.
3. Correct - let a human or coding agent make a scoped change.
4. Verify - rerun the same command and compare the evidence.

No LLM is used in the deterministic CI enforcement path.

Memi InterfaceBench

Memi InterfaceBench is the release benchmark for professional product design, frontend engineering, and native design engineering. It measures whether an agent can understand an existing product, make a scoped interface change, preserve its system, and prove the result. The benchmark is currently a specification with results: null, not a published Memi score.

Benchmark contract	Value
Target corpus	100 target tasks
Evaluation split	20 public development, 60 private test, 20 rolling holdout
Coverage	10 task families across React Web, React Native/Expo, SwiftUI, desktop canvas, and cross-platform work
Initial reproducible slice	5 pinned seed tasks, presented as Examples 1, 2, 3, 5, and 7
Repetitions	At least 3 independent paired runs per condition
Quality	75% work-product score and 25% source-trust score after hard gates
Efficiency	Measured cost, latency, tokens, retries, and interventions; tool-call count is diagnostic only
Release rule	Quality non-inferiority plus positive cost and latency evidence with 95% confidence

The work-product score covers functional correctness (22%), visual fidelity and craft (14%), product judgment (12%), interaction and state quality (10%), accessibility (10%), responsive behavior (9%), design-system integrity (9%), maintainability (7%), performance (4%), and platform fidelity (3%). Source trust separately scores evidence provenance (30%), fixture integrity (25%), decision trace (20%), source alignment (15%), and uncertainty honesty (10%).

The protocol adapts the professional-work-unit and bespoke-rubric approach from Harvey BigLaw Bench, pinned real-repository tasks from SWE-bench Verified, rendered visual evaluation from Design2Code, functional journeys from WebArena, and blinded pairwise preference from WebDev Arena. Accessibility and performance gates follow WCAG 2.2 and Core Web Vitals.

Measured 2.7 candidate evidence

These results predate the full 100-task InterfaceBench run. They are reported as release evidence, not converted into an InterfaceBench score.

Evidence set	Scope	Quality guardrail	Token savings	Wall-time savings	Decision
Writable product workflows	5 paired, multi-minute E2E runs	Automated acceptance passed in every case; professional quality unscored	0.55% mean, 95% CI -51.23% to 37.30%	5.66% mean, 95% CI -24.30% to 25.86%	Claim not verified
Six-repository read-only suite	6 paired runs	100% pass rate in both conditions	19.92% mean, 95% CI -0.23% to 40.07%	0.50% mean, 95% CI -11.72% to 14.55%	Claim not verified
Latest Example 2 Router v2 calibration	1 paired Expo run	Automated acceptance passed in both conditions	14.15%	25.08%	Calibration only

Published 2.7.3 verification records:

• 2,187/2,187 tests passing across 304 files locally and across the Linux, macOS, and Windows Node 20/22/24 release matrix.
• 70.57% statements, 58.87% branches, 79.82% functions, and 72.35% lines. This is a measured improvement from 64.91%, 53.73%, 73.11%, and 66.62%, respectively, but remains below the 80% repository target.
• Zero production dependency vulnerabilities.
• The published 54-file npm artifact measures exactly 577,370 bytes compressed and 1,990,260 bytes unpacked, within the 750,000-byte and 2 MiB release budgets.
• No defensible per-run USD cost from the available subscription traces, so total tokens are explicitly labeled as the cost proxy.

Historical qualityScore: 100 workflow records mean automated acceptance passed, not senior-practitioner quality. New records label that evidence, cap it at 80, and reserve higher scores for calibrated practitioner review.

The greater-than-25% claim remains not verified. Inspect the LaTeX research paper, technical release brief, reproducible paper source, writable workflow report, six-repository report, Example 2 tool profile, and the complete InterfaceBench contract.

Memi DesignWorkBench v2

Memi DesignWorkBench v2 expands the quality benchmark from interface implementation to complete professional design practice. The frozen foundation contains 300 task contracts across product, graphic, motion, spatial/XR, agentic, iOS, Expo, Android, native kits, design systems, design engineering, service, content, interaction, and research work.

The corpus is split into 60 public development tasks, 180 private test tasks, and 60 rolling holdouts. Acceptance, professional quality, source trust, and efficiency are separate outputs. Track scores use a geometric composite so a strong design-system result cannot conceal an unsupported native or spatial track.

The foundation is not a release score. Real fixtures, device runners, and practitioner calibration remain mandatory. Synthetic practitioner ratings are prohibited, and the release gate fails until every claimed track has at least four qualified practitioners, five artifacts per practitioner, two blinded ratings per artifact, reliable graders, and measured private plus holdout results. See the current readiness report and execution goal. Three of eight runner profiles now have reproducible, hash-bound evidence: browser/Playwright, motion rendering, and artifact/provenance validation. All 60 public development tasks now also have deterministic, hash-bound candidate fixture packs. Those packs are explicitly benchmark-authored synthetic material: 60 prepared, 0 independently verified. The remaining runners, all 300 verified fixtures, and all practitioner results remain blocked rather than inferred.

DesignWorkBench certification is separate from package publication: check:release verifies integrity and readiness, while check:certification requires external practitioners, private tests, and holdouts.

See the ORCID-ready benchmark paper and its unexecuted cross-harness protocol.

Choose your integration

Surface	Start here	Best for
One-time CLI audit	npx -y @memi-design/cli@2.7.4 diagnose . --no-write	Trying Memi without installing
Global CLI	npm i -g @memi-design/cli	Daily local use
Agent Skill	npx skills add memi-design/memi --skill audit-frontend-design	Codex, Claude, Cursor, and compatible agents
GitHub Action	uses: memi-design/memi@8aa4649f412bbcaaf2af4ee209bf79016566f035	Pull-request design CI
MCP server	memi mcp start --no-figma	Any MCP client
Studio	brew install --cask memi-design/memi/memi-studio	Supervised macOS workflows

The CLI and focused skills are primary. Studio, Figma, research, scaffolding, registries, and the larger tool router are deeper paths.

Design CI

Pin the release commit so every pull request runs the same code:

name: design
on: [pull_request]

permissions:
  contents: read

jobs:
  memi:
    runs-on: ubuntu-latest
    permissions:
      contents: read
      security-events: write
    steps:
      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
        with:
          fetch-depth: 0
      - uses: memi-design/memi@8aa4649f412bbcaaf2af4ee209bf79016566f035 # v2.7.4
        with:
          version: "2.7.4"
          report: true
          upload-sarif: true

The Action adds code-scanning annotations, a step summary, and a memi-design-health artifact. Existing debt can be baselined while new debt fails the gate.

GitHub Action guide · CI recipes · team rollout

Agent and MCP setup

Install a complete project kit:

memi agent install codex --project .
memi agent install claude-code --project .
memi agent install cursor --project .
memi agent install grok-build --project .

Start the local MCP server without Figma:

memi mcp start --no-figma

{
  "mcpServers": {
    "memoire": {
      "command": "memi",
      "args": ["mcp", "start", "--no-figma"]
    }
  }
}

Codex plugin marketplace:

codex plugin marketplace add memi-design/memi --ref main --sparse .agents/plugins --sparse plugins/memoire

Agent stack guide · copy-paste recipes · full skill router

Proof you can inspect

• design-sandbox - runnable Next.js, Tailwind, shadcn, MCP, and Agent Skills integration.
• Release gates - package, provenance, clean install, MCP, plugin, binary, and public-surface checks.
• Current release truth - one source for npm, GitHub, Action, Studio, and website versions.
• Reproducible case studies - pinned evidence, abstentions, and paired benchmark protocols.
• Audit reports - timestamped findings, evidence gaps, score caps, and owners.
• llms.txt - compact machine-readable product map.

Trust defaults

• Read-only audit by default
• No npm install-time lifecycle scripts
• No account or API key for the first audit
• No source upload or covert telemetry
• Figma installation and connection are explicit
• Agent kit writes support --dry-run --json
• Immutable Action pins and release provenance
• MIT core with third-party boundaries documented in NOTICE

Community

• Ask a question
• Show what Memi found
• Report a bug
• Request a feature
• Contribute

Useful contributions include reproducible audit fixtures, framework adapters, skill improvements, accessible UI cases, shader and motion checks, and real before/after reports.

More commands

CLI reference

Command	What it does
memi diagnose [target]	App-quality graph and file-anchored findings
memi ux audit [target]	UX tenets and product-state risks
memi craft audit [target]	Hierarchy, rhythm, convention, and responsive critique
memi tokens --from <path>	Extract tokens, aliases, modes, and drift
memi agent brief . --json	Compact evidence contract before UI work
memi scaffold component <Name> --json	Dry-run Atomic Design scaffold
memi ios brief --json	Apple-platform design brief
memi shadcn export --out public/r	Export a shadcn-compatible registry
memi research design --json	Research-to-design package
memi ci	Baseline-aware pull-request gate
memi mcp start --no-figma	Local MCP server

Install without npm

curl -fsSL https://memoire.cv/install.sh | sh
brew install memi-design/memi/memoire
docker run --rm -it -v "$PWD:/work" -w /work ghcr.io/memi-design/memi --help

License

Studio interface references and adapted components include Hermes WebUI and the MIT Warp UI framework boundary around warpui_core and warpui; Warp AGPL application and client code is not copied into Memi.

MIT. See NOTICE for optional adapters and complete third-party attribution.