# Visual Memory MCP [Health: Active]

**Category:** 🧠 Knowledge & Memory  
**Repository:** https://github.com/putervision/vision-memory-mcp.git  
**GitHub Stars:** 40  
**npm Downloads (last month):** 4643  
**Views:** 1  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/visual-memory-mcp

## Description
Local visual UI cache for AI agents using perceptual hashing + CLIP to cut vision token use.

## Claude Desktop Quick Installation
Install path detected from listing signals. Uses `npx` (confidence: high):

```json
"mcpServers": {
  "visual-memory-mcp": {
    "command": "npx",
    "args": ["-y","@putervision/vision-memory-mcp"]
  }
}
```

## Documentation

## What the Visual Memory MCP MCP server does

Visual Memory MCP MCP server stores and retrieves visual UI state for AI coding assistants. It is intended for workflows where an agent repeatedly receives screenshots, screen recordings, or related interface states and would otherwise spend vision-model tokens reanalyzing them.

The server can analyze screenshots, search remembered states by text or image similarity, identify likely interface elements, and retain transitions between states. It also supports visual specification baselines, named snapshots, state comparison, video-derived timelines, and deletion of stored states. The repository describes local LanceDB storage, local CLIP embeddings, perceptual hashing, and PII redaction as part of its local-first design.

## How it works

Visual Memory MCP MCP server uses a four-stage lookup path. It first checks an in-memory cache, then scans perceptual hashes, then searches local CLIP vectors. If those checks miss, the architecture falls back to a vision LLM, ingests the resulting state, and saves it for later retrieval. Cached results can include descriptions and grounded interface elements.

The server represents UI changes as state transitions. Agents can record action outcomes, find navigation paths, wait for a target visual state, and compare layouts or recorded trajectories. WebM and MP4 recordings can be converted into searchable keyframe timelines.

## Setup and configuration

Node.js 18.18.0 or newer is required. Install the package globally with npm, run `vision-memory-mcp init --yes` from a project root, and add a stdio entry that invokes `vision-memory-mcp run` to the MCP client configuration. Initialization scaffolds database directories, a `.gitignore`, `.env`, and IDE rules.

Use `vision-memory-mcp run --skip-model-load` when heavy CLIP model downloads should be skipped, such as in an offline or air-gapped environment. The CLI also includes `doctor` and `doctor-global` checks, plus commands for inspecting stored states, registering visual specifications, saving snapshots, ingesting recordings, and opening the graph viewer.

## Tools and capabilities

The 15 consolidated MCP tools cover these areas:

- Screenshot analysis with perceptual hashing and accessibility-tree parsing.
- Text and image semantic retrieval through `recall_memory`.
- Session summaries, cache metrics, and compact context export.
- CSS-selector and coordinate prediction for element grounding.
- Action outcomes, blockers, navigation paths, and visual-state polling.
- Video ingestion, timeline search, state comparison, and trajectory export.
- Visual specification contracts, snapshots, state undo, and privacy-focused state deletion.
- Evidence-pack creation for linking visual records to task-memory data.

Visual Memory MCP MCP server is therefore a fit for frontend agents, end-to-end test recordings, visual regression checks, and workflows that need compact visual context rather than repeated full image analysis.

## Limitations and notes

Its standard configuration uses a local stdio process, local vector storage, and local CLIP processing. Skipping model loading changes availability of the CLIP stage and may affect retrieval behavior.

The README names Cursor, Claude Code, Gemini, Copilot, Claude Desktop, Antigravity, Windsurf, Zed, and Roo Code in its integration material. Only clients supported by the directory schema can be represented in the listing metadata.

_Full upstream README: https://allmcps.com/mcp/visual-memory-mcp/readme_