# vision-memory-mcp

A local MCP server that caches screenshots and UI states using perceptual hashing and vision embeddings, so a coding agent recognizes a screen it has already seen instead of paying for a fresh vision-model call each time.

vision-memory-mcp is a zero-infrastructure, local-first Model Context Protocol server and command-line tool that gives AI coding assistants such as Cursor, Claude Code, Gemini or Copilot a cache of visual state. It stores perceptual hashes, local CLIP embeddings and accessibility trees of screens it has seen before, along with the transitions between them, so a repeated layout is recognized from the cache in a few milliseconds instead of triggering another call to a vision-capable model.

It can also ingest WebM and MP4 recordings, such as end-to-end test runs or screen captures, turning them into searchable keyframe states and transition graphs, and it can map an interface element to its CSS selector and coordinates for deterministic interaction. Registering a visual baseline lets later runs be checked against it for visual regressions, and a companion state-memory package can be bridged in to combine task tracking with this visual memory.

It installs as a global npm package and is initialized once per project.

Website: https://visionmemorymcp.com