A local media indexer and natural-language search engine that helps AI agents organize, describe, and query photo and video libraries while optimizing GPU memory.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install mneme-vision
2. Manual Installation
Copy the skill folder to one of these locations
~/.openclaw/skills/ <project>/skills/ Priority: Workspace > Local > Bundled
3. Prompt Installation
Copy this prompt to OpenClaw to install it automatically.
Help me install mneme-vision using Clawhub. If Clawhub is not installed, install it first (npm i -g clawhub).
Prefer to download?
Get the raw skill files in a ZIP archive.
What is Visual Memory?
The Visual Memory skill empowers local AI agents to seamlessly catalog, search, and manage rich media assets like photos, videos, reels, and audio clips. By utilizing advanced local vector indexing and visual language models (VLMs), this workspace addition transforms raw directories of creative content into searchable semantic databases. It runs entirely locally to preserve privacy and integrates with tools like mcporter and resident models (such as Ollama).
Unlike traditional cloud-based digital asset management, this capability operates inside your local environment as part of the broader ecosystem of Openclaw Skills. Developers and creators can query media libraries using natural language, perform reverse image searches, and inspect specific assets using vision models. It also includes sophisticated VRAM and GPU resource orchestration, ensuring your local models cooperate nicely during heavy workflows.
Visual Memory Use Cases
- Finding specific shots, finished videos, or archive clips using natural-language search descriptions.
- Performing reverse image searches (
media_search_by_image) to find visually similar assets in massive local folders. - Auto-generating detailed descriptions of image and video contents via local visual language models (
media_describe). - Indexing uncataloged local creator libraries (
media_index) to convert directory files into structured search results. - Dynamic GPU memory management (
gpu_release,gpu_reclaim) to optimize VRAM usage when switching between video editing, image generation, and LLM text workflows.
How Visual Memory Works
- Attachment or Directory Scan: The AI agent evaluates a given path or temporary media attachment.
- Local Indexing: If needed, the agent calls
media_indexto leverage mcporter and local indexing capabilities to digest folders without uploading data to the cloud. - Semantic and Visual Querying: Users query their media library via natural language (
media_search) or image-to-image similarity search (media_search_by_image). - GPU VRAM Negotiation: Before running heavy local model operations (like VLM tagging or upscale), the agent coordinates VRAM by calling
gpu_statusand releasing or reclaiming Ollama model space usinggpu_releaseandgpu_reclaim. - Structured Rich Output rendering: The agent outputs findings using the
media_artifacts.v1schema, displaying previewable cards, titles, paths, and reasons inside the host application interface.
Visual Memory Setup
Prerequisites
This skill requires the local binary mcporter to perform indexing and visual metadata extraction. Ensure it is installed on your local path.
Quick Start
Enable the Visual Memory skill within your configuration for Openclaw Skills:
# Ensure mcporter is available globally
mcporter --version
# Configure the skill in your local workspace setup
openclaw install visual-memory
No cloud API keys are required as all search and processing operations happen on your local machine.
Visual Memory Data Schema & Taxonomy
The Visual Memory skill structures results using the neutral media_artifacts.v1 JSON specification to preserve maximum compatibility with host UI components.
Return Schema Fields
| Field | Type | Description |
|---|---|---|
title |
String | Filename or human-readable clip name |
kind |
String | Asset classification: image, video, audio, or unknown |
path |
String | Absolute local path, kept available for copy and reveal actions |
reason |
String | A brief, natural-language explanation of why this media matched the query |
timestamp | time_range |
String/Object | Specific temporal coordinates (when available for video or audio) |
preview |
Boolean | Set to true when the file format can be natively rendered in the host UI |
Temporary session media (such as direct user uploads) are treated as transient context and are not committed to the index database unless explicit instructions are given.
Visual Memory Advanced Features
- Advanced VRAM Lease Coordination: Seamlessly coordinate VRAM with local generation tasks using
gpu_status,gpu_release, andgpu_reclaimtokens to prevent out-of-memory errors on consumer GPUs. - Zero-Cloud Local Privacy: All indexing, classification, and vector search operations happen strictly on-device, satisfying strict privacy policies for production-sensitive asset libraries.
- Semantic Visual Search: Search your library using abstract descriptions like "cinematic overhead shot of a rainy street" without manual tag setup.
SKILL.md
Loading
Related Openclaw Skills
Veteran Proxy Client
A cross-platform CLI proxy client that manages subscription updates, multi-node switching, SOCKS5/HTTP proxies, and advanced split routing rules.
World Order
An executable geopolitical analysis toolkit modeled on Henry Kissinger's World Order framework to evaluate balance of power, diplomatic strategy, and international security.
Who Gets to Be Indian (Dina Gilio-Whitaker)
An executable educational toolkit designed to help users navigate the socio-political complexities of Native American identity, sovereignty, and ethnic fraud.
Surviving Climate Anxiety
An executable AI agent toolkit for coping with climate-related distress, building emotional resilience, and taking sustainable environmental action without burnout.
Work Rules! (Google People Operations)
An executable AI skill based on Laszlo Bock's Work Rules! to automate and guide team management, high-performance hiring, and culture-building using Google's proven people operations framework.
Winning
An executable leadership toolkit designed for AI agents to deploy Jack Welch's performance-driven management principles directly into corporate workflows.
Featured*
