memi

memi is an AI design workbench built specifically for product design teams who want to keep their design decisions, system context, and agent runs in one cohesive environment. Unlike generic AI chat tools, memi is a signed macOS application that bundles Codex and Claude Code agents with persistent project memory, Figma handoff capabilities, and a structured receipt system for every agent action. Its core value is giving designers a single workbench where design memory stays readable, traceable, and reusable across sessions. The tool recognizes that modern design work involves running AI agents alongside human decisions, and it provides a harness that connects these threads without losing context. For teams managing design systems, user research, or complex UI audits, memi offers a purpose-built runtime that keeps the product system understandable and actionable.

The main pain point memi addresses is the ephemeral nature of AI design conversations. When designers use ChatGPT, Codex, or Claude directly, the context, decisions, and artifacts generated in one session are rarely carried over to the next. This leads to repeated setup, lost insights, and a fragmented design process. memi solves this by capturing every agent run as a structured receipt and storing design memory as editable Markdown or YAML. It also surfaces design system context—like tokens, components, and screenshots—that agents can read and reuse. Without this, teams waste time re-explaining their design system, re-running audits, and re-capturing screenshots. With memi, the work accumulates into a readable project memory layer that agents and humans can both depend on.

The first major feature group is Design Memory, which functions as a structured project memory layer. Instead of hiding context inside opaque prompt soup, memi stores research, specs, tokens, decisions, and reviews in plain Markdown and YAML files. Designers can edit these files directly, and agents can read, diff, and reuse them across runs. This means every design decision is captured as an artifact that can be audited, reviewed, or fed back into future agent runs. The memory is not a black box—it's inspectable and versionable, like code. When a designer runs an audit or creates a component, the output is saved as structured memory that grows with the project. This feature eliminates the need to manually transfer notes between tools or re-explain context to AI agents.

The second major feature group is Agent Runs with Receipts, which orchestrates Codex and Claude Code sessions with explicit controls. Designers launch an agent from the memi workbench, choosing the model, permissions, and session parameters. As the agent runs, the 'run spine' displays prompts, plans, tools used, files accessed, and cost in real time. Every action is recorded as a receipt that remains compact until the user needs the raw log. This transparency is crucial for design teams that need audit trails—knowing exactly what an agent saw, what it changed, and how much it cost. Designers can also cancel runs safely and save results for later review. This feature turns agent use from a black-box experiment into a documented workflow.