memi

[Memi — the design layer for agentic AI.]

[npm version] [weekly npm downloads] [CI] [GitHub stars] [MIT license]

Memi

The design layer for agentic AI.

Give your coding agent an interface brief before it edits. Memi maps the UI already in your repository, surfaces file-anchored accessibility and design-system risks, and gives you a deterministic check to rerun before merge. Start with the CLI, then add the same gate to every pull request.

Memi Studio is available today; Memi Canvas is currently in development. No account, API key, Figma file, global install, or daemon is required for the first audit.

Start with your next interface · Get Memi Studio · Read the research

Quickstart: find your first interface issue

Run one non-destructive audit in any frontend repository. It needs no account, API key, Figma file, global install, or daemon.

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

The result carries normalized finding IDs, confidence, provenance, and file:line evidence so an agent can act on a specific finding instead of guessing.

Give the same context 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 changes that will matter most to users, reuse the existing system, and verify the result after the patch.

If Memi catches a real interface issue in your project, share the finding. Real reports are the most useful signal for what to improve next.

Cost and offline operation

Local source diagnosis uses deterministic rules: no model call, API key, or model fee. It checks the repository you provide and returns findings your agent can verify. Installation, CPU time, CI runners, and any agent you choose still have costs. The published research has not established an end-to-end dollar or token savings advantage.

Start with one diagnosis, address the highest-confidence finding, and rerun the same check. Use --agent-context for a bounded file index instead of sending a whole repository to a model. Its routing is heuristic; a smaller context is not proof of a cheaper successful task. --files scopes reported findings without launching Git; it still scans the tree for aggregate statistics.

Published beta: Install the exact 2.8.0-beta.2 package for the 2.8 frontend workflow. Stable remains 2.7.9; the beta is not a stable release. See current release state and known limitations.

The beta2 npm release record identifies its published source and immutable artifact evidence. The beta1 npm release record preserves the earlier release's history; it does not validate beta2. The 2.8 locked default and explicit capability contract do not apply retroactively to 2.7.9.

npx -y @memi-design/[email protected] --versionnpx -y @memi-design/[email protected] agent brief . --frontend --intent "Improve this interface" --json

The beta keeps managed security review, SwiftUI parity, and full DesignWorkbench certification explicitly pending. It is not stable or employer-approved. Publication provenance and platform receipts belong to the exact released artifact.

The beta prevents local diagnosis from writing project files by default, keeps paid integrations optional, and requires explicit grants for networking, subprocesses, and source report persistence. An npm install needs registry access; offline first-run claims require the separately verified bundle. See the acceptance ledger, known limitations, and release truth.

The 2.8 frontend workflow: the beta provides a repository-aware frontend brief, four useful locked MCP tools, validated Figma/Paper evidence inputs, and explicit component-reuse conflicts. See the frontend workflow for runnable beta commands and the release plan for remaining gates. Managed independent security scanning and the reviewed-candidate-audit and swiftui-rendered-rerun receipts remain pending for stable. Paper pixel parity is unassessed. Of 158 inventoried CLI paths, 105 remain deferred; broad grants do not enable them. No complete-task dollar or token savings claim is established.

What 2.8 changes for frontend agents

The practical benefit is more useful local evidence and clearer failure signals: a brief identifies an existing component before an agent duplicates it, diagnosis distinguishes unchecked categories from assessed findings, and diagnose --receipt-only provides a separate metadata-only result without persisting a source report. Regression tests also exposed and corrected candidate defects in bounded file reads, malformed event handling, and failure exit codes. Those corrections improve specific behavior; they do not establish universal design quality, security certification, or lower task cost.

Earlier packed results at source 2a6d0e44 are historical evidence for those exact bytes. Later defensive runtime fixes supersede that artifact as a release candidate; each corrected source checkpoint is verified through a separate local package and digest. See the verification record for the source, artifact, and scope of each result. Local candidate receipts remain evidence for their original bytes; they do not replace the published npm record.

The repository remains authoritative for code. Memi supplies evidence; Codex edits and runs the project's actual checks. Existing custom CSS and component systems do not need a shadcn migration. This beta deliberately defers many legacy command paths; it is not yet a drop-in upgrade for every 2.7 workflow.

Put the check on every pull request

Copy examples/github-actions/memi-design.yml into your repository as .github/workflows/memi-design.yml. The starter is pinned to the reviewed public Action commit and gives reviewers:

• a PR check that fails only on newly introduced interface debt;
• a memi-design-health artifact with the human-readable report; and
• SARIF annotations when the repository grants security-events: write.

The workflow does not need an API key or a Memi secret. Fork pull requests still receive the check and report; SARIF upload is skipped automatically when GitHub does not grant that permission.

If you prefer to configure it by hand, the complete GitHub Action guide documents every input, output, permission, and evidence file.

One product layer, three surfaces

See the product

What Memi adds to an agent workflow

Compatible with the shadcn registry and v0 design systems.

Evidence at a glance

The V15 confirmatory audit is a public technical disclosure, not a leaderboard. It separates receipt admission, rendered design quality, functional acceptance, and resource observations.

Separate historical release record: the 2.7 candidate record reported 2,187 / 2,187 tests passed. It is release evidence, not part of V15 and not proof that every project benefits.

Benchmarks and paper

[V15 benchmark preview: scoped quality non-inferiority passed for two task families; the study makes no general superiority, speed, or cost claim.]

Quality non-inferiority passed for the scoped Buzzr and Paraform task families. The full paper reports exclusions, failed paths, and limitations without imputation. No superiority, speed, or dollar-savings claim is made. Read the conference-style audit PDF, inspect the protocol and receipts, or review the V17 preregistration.

Memi InterfaceBench v1 is a 100 target tasks specification with 5 pinned seed tasks; it is not an aggregate performance score. The historical candidate record reported 2,187/2,187 tests and 70.57% statements coverage. The greater-than-25% claim remains not verified. Inspect the benchmark contract and workflow evidence.

Memi DesignWorkBench v2 holds 300 task contracts and requires practitioner calibration before any certification claim.

Prompts that map to real workflows

Research, stated plainly

The research is disclosure material, not a product leaderboard. It keeps functional, rendered-quality, and resource evidence separate so a result cannot be made to say more than the study supports.

Choose your integration

GitHub Action

name: designon: [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@5fcbf39e1255af0c14c5a17ba6bde8cf1206e525 # v2.7.9        with:          version: "2.7.9"          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 newly introduced debt fails the gate.

GitHub Action guide · CI recipes · current versions

Agent and MCP setup

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

{  "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

Trust and proof

• Release gates — package, provenance, clean-install, MCP, plugin, binary, and public-surface checks.
• Current release truth — the public versions for CLI, Studio, and website.
• Reproducible case studies — pinned evidence, abstentions, and paired protocols.
• Dependency trust ledger — direct dependency purpose, dynamic boundaries, and review policy.
• Trust Core index — locked-default contract and evidence status for the 2.8 beta.
• Threat model and egress map — protected assets, side-effect boundaries, and destinations.
• Data retention and uninstall/recovery — what can persist and how to preserve state.
• Dependency/license review and supported platforms — artifact and sandbox gates.
• Employer review packet — artifact-bound approval checklist; internal use requires written employer approval.
• Known limitations and release truth — beta versus stable language and open evidence.
• Organization compatibility — sibling surfaces, including independently gated Studio and Canvas.
• llms.txt — compact machine-readable product map.

The deterministic first audit requires no source upload or telemetry. Connected integrations expand the boundary and require explicit review. Memi also has no npm install-time lifecycle scripts, provides agent-kit --dry-run --json, pins the Action immutably, and documents third-party boundaries in NOTICE.

Community

We welcome contributions. See CONTRIBUTING.md for setup and pull-request guidance. Bugs and feature requests belong in issues; questions and real project reports belong in Discussions.

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

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.