Openclaw mneme vision

Use local visual memory tools to index and search photos and videos from creator media libraries.

Visual Memory

Use this skill when the user wants to index, search, organize, or inspect local photo and video libraries.

When the user provides an uploaded photo/video/reel as an attachment, treat it as temporary session context first. Use its local path to answer the current question or run visual similarity search when appropriate. Do not call media_index or otherwise add the attachment to durable visual memory unless the user explicitly asks to add, save, index, catalog, or remember it.

Prefer media_index when the user points at a folder that has not been indexed. Prefer media_search for natural-language searches such as finding shots, finished videos, thumbnails, archive images, or clips matching a description. Prefer media_search_by_image when the user provides an image and wants visually similar items. Use media_describe when the user asks what is in a specific media file.

Use gpu_status before GPU-heavy local tasks when the user is coordinating video work, image generation, upscaling, VLM tagging, or another local model process. Use gpu_release to unload resident Ollama models and hand VRAM to that workflow, and use gpu_reclaim with the returned token when the workflow is done. Use gpu_evacuate only when the user asks to free VRAM immediately without creating a lease.

Return media results as previewable content cards when the host supports rich output. Each result should include:

• title: filename or human-readable clip name.
• kind: image, video, audio, or unknown.
• path: absolute local path, kept available for copy/reveal actions.
• reason: short explanation for why this media matched.
• timestamp or time_range: when available for video/audio hits.
• preview: true when the file can be rendered by the host UI.

Mneme returns neutral media_artifacts.v1 JSON. Preserve the host application's native styling; do not introduce custom colors, cards, gradients, or branding unless the host already provides that UI.

Do not dump long bare paths as the main answer. Put the path behind a Copy path or equivalent action when rich output is available, and show the user the actual image/video/audio preview first. When rich output is not available, use concise Markdown links or a compact table with path values.

Do not invent matches when the tool returns none.

Do not upload local media to a cloud service unless the user explicitly asks for that separate action.