Meta:media Skill for Claude

/media-memory — Multimodal Memory System

You have access to a persistent multimodal memory system at ~/.claude/media-memory/. It stores every piece of media (images, video, audio, files) with rich metadata and Gemini Embedding 2 vectors in ChromaDB.

Directory Layout

~/.claude/media-memory/
  assets/          # stored media files
  chroma/          # ChromaDB vector store
  metadata.db      # SQLite structured metadata
  scripts/
    ingest.py      # ingestion + embedding
    search.py      # search with filters
    schema.py      # metadata models

Commands

All commands run from ~/.claude/media-memory/ using uv run.

Ingest (store + embed)

cd ~/.claude/media-memory && uv run scripts/ingest.py "<file_path>" \
  --source "user|generated|url|ingested" \
  --description "Natural language description of the media" \
  --tags "tag1,tag2,tag3" \
  --type "image|video|audio|document|file" \
  --text "Extracted text or transcript content"

Search (hybrid: semantic + metadata)

cd ~/.claude/media-memory && uv run scripts/search.py "search query" \
  --type image \
  --source user \
  --tags "architecture,diagram" \
  --from "2026-03-01" \
  --to "2026-03-28" \
  --limit 10 \
  --mode hybrid|semantic|metadata \
  --json

Recent items

cd ~/.claude/media-memory && uv run scripts/search.py --recent --limit 10

Stats

cd ~/.claude/media-memory && uv run scripts/search.py --stats

Behavior Rules

On Ingest (when user sends or generates media)

• Copy the file to assets/ via ingest.py
• ALWAYS provide --description with a rich natural language description of the content
• ALWAYS provide relevant --tags for semantic categorization
• Set --source accurately: user (user sent it), generated (Claude/AI created it), url (downloaded), ingested (bulk import)
• For screenshots: describe what's visible (UI elements, text, code, diagrams)
• For documents: extract key text into --text
• Report the result to the user: "Saved to media memory: {description}"

On Search (when user asks about past media)

• Use --mode hybrid by default (combines semantic + metadata)
• Add --type filter when user specifies media kind
• Add --tags filter when user mentions categories
• Add date filters when user references timeframes ("last week", "this month")
• Show results with descriptions and asset paths
• Offer to open/display the asset if it's an image

Proactive Recall

When a conversation topic overlaps with stored media:

• Run a quick semantic search with the current topic
• If relevant results found (similarity > 0.7), mention: "I found a related {type} in media memory: {description}"
• Don't be noisy — only surface genuinely relevant assets

Environment

• No API key needed — uses ChromaDB's built-in local embeddings (all-MiniLM-L6-v2 via onnxruntime)
• Everything runs locally, zero external calls
• ChromaDB: local persistent storage, cosine similarity
• Model cached at ~/.cache/chroma/onnx_models/ (downloaded once on first use)

Metadata Schema

| Field | Type | Description |

|-------|------|-------------|

| id | string | Auto-generated: {type}_{hash}_{stem} |

| filename | string | Original filename |

| type | string | image, video, audio, document, file |

| timestamp | ISO 8601 | When ingested |

| source | string | user, generated, url, ingested |

| description | string | Natural language description |

| extracted_text | string | OCR / transcript / content |

| tags | JSON array | Semantic tags |

| original_path | string | Where it came from |

| asset_path | string | Path in assets/ |

| embedded | boolean | Whether vector is in ChromaDB |