Bug/Docs Issue: “Multimodal (text/images/audio/video)” claim doesn’t match main-branch implementation

[ved015]

Summary

The README/paper/docs describe MIRIX as supporting multimodal inputs (text, images, audio, video). In the current main branch (v0.1.6), the code path for the memory layer appears to natively support text + images, but not audio/video as first-class inputs.

What I expected

• Ability to send audio/video inputs (e.g., as message content types) and have MIRIX:
  • ingest them,
  • optionally transcribe/extract representations,
  • pass them to the LLM (or convert to text/frames),
  • and then store results in memory.

What actually happens

• Message content types include text, image_url, file_uri, google_cloud_file_uri (no audio / video).
  • Code: mirix/schemas/mirix_message_content.py
• Images are handled end-to-end: they become ImageContent(image_id=...) and are expanded into real image bytes/URL for provider requests.
  • Code: mirix/utils.py (convert_message_to_mirix_message)
  • Code: mirix/llm_api/openai_client.py and mirix/llm_api/anthropic_client.py (fill_image_content_in_messages)
• Audio/video are not handled as native content types and are not automatically transcribed/processed in the main /memory/add pipeline.
  • There is a transcription helper (mirix/voice_utils.py), but it does not appear wired into the REST/queue processing path.
• Additionally, OpenAI provider code explicitly does not support file_uri as LLM input.
  • Code: mirix/llm_api/openai_client.py (raises NotImplementedError for file_uri)

@wangyu-ustc