# Media

Common images, documents, audio, voice, video, and animations are received
and sent. Support means: durable content-addressed storage, authenticated
display, and automatic understanding on triggering use.

- Inbound media downloads stream through the service API into the CAS store;
  chat rows reference digests, never blobs.
- Analysis happens when a message triggers a wake or recall selects media,
  not on every group message. Results cache by digest + capability + model +
  schema version.
- Curated ZAI tools do the work (see the zai-capabilities skill); provider
  limits are explicit (audio <=25 MB/30 s, images <=10 MB, PDF <=50 MB/30
  pages, MCP video <=8 MB). Unsupported, corrupt, encrypted, or oversized
  input fails with a clear message; will never claims false understanding.
- Generated images/videos are stored and sent as digests.
- Retained forever with the conversation; hidden messages tombstone in UI
  and recall, bytes stay in archives.