Qwen Omni Memory

Omni Memory is one of the Qwen team's Qwen-MM-Plugins. It turns a long video into an audio-visual memory your agent can question. A Qwen Omni model reads each 30-second window together with its audio, so every spoken line is tied to the person saying it, how they said it and the sounds around it. What you can do with it - Ask about a short clip (under about 10 minutes) and get an answer from one watch, with no memory built - Build a memory of a long video, or of several files from one recording session, then ask follow-up questions without re-watching it - Find everything one person said, or who said something, with timestamps and tone - Look up who appears, the facts and relationships the video establishes, and what happens at a given time - Re-watch a few stored clips with their audio when the memory is missing a detail How it connects It's a Claude Code plugin with a skill and a local MCP server that runs with uvx. In Claude Code, run claude plugin marketplace add https://github.com/QwenLM/Qwen-MM-Plugins.git, then claude plugin install qwen-mm-plugins-omni-memory@qwen-mm-plugins. Qwen's guided installer also sets it up in Codex, OpenClaw, Qwen Code, Gemini CLI and other agents. Good to know - It needs a DashScope API key (DASHSCOPE_API_KEY), set in your environment or in ~/.qwen-mm-plugins/config through Qwen's installer, plus ffmpeg to build a memory. Use the config file when your agent is launched from a desktop app rather than a shell. - Building a memory makes one Omni model call per 30 seconds of video. It's paid for once, and the memory then answers later questions. - Video goes to DashScope for the Omni model to read; on the official endpoint, oversized files are uploaded to temporary storage there. - Memories are saved next to the video, in a .memory folder.