Assistant Profile and Preferences
Set a default assistant style and save your own preferences through conversation. They are stored separately, and upgrades do not overwrite your local profile.
Set Preferences in Conversation
- “Call me Captain from now on.”
- “Keep your replies concise from now on.”
- “When we talk, your name is River.”
Explicit long-term requests become user preferences. “Keep this reply short” is temporary and should not become a lasting setting.
Edit the Default Profile
On first start, a template creates <config-dir>/ASSISTANT.md. Edit it to change the default name, personality, relationship, and communication style. Changes apply to the next voice session.
To use another file:
QWEN_AUDIO_AGENT_ASSISTANT_PROFILE_PATH=/absolute/path/to/ASSISTANT.mdThe profile does not select the speaking voice; configure that under voice frontend settings. It cannot override tool permissions, task routing, or core runtime rules.
Where Settings Live
| Content | Default file | Example |
|---|---|---|
| Default assistant profile | <config-dir>/ASSISTANT.md | Default name and style |
| Explicit long-term preferences | <data-dir>/USER.md | “Call me Captain” |
| Long-term facts and decisions | <data-dir>/MEMORY.md | “I live in Hangzhou” |
The current request takes precedence over saved preferences; preferences can override the default profile. Facts support understanding, not permission grants or execution instructions.
Changes made through conversation apply immediately. Direct file edits apply in the next voice session. Do not store passwords, keys, or verification codes here. See data directories.
Optional Preference Learning
By default, preferences come from explicit requests. Set QWEN_AUDIO_PREFERENCE_LEARNING=on to let the default memory provider observe a few characteristics after sessions and write cross-session confirmations into the inferred section of USER.md.
This is off by default and makes additional text-model calls. Explicit preferences take precedence over inferences. You can delete unwanted inferences from the file. See Preference Learning for thresholds and implementation details.
Privacy and Other Providers
Files stay on the Gateway host, but relevant profile, preference, and memory content is sent as context when using cloud models. Local storage does not mean fully offline processing.
See Memory for automatic extraction, inspection, and deletion. Use VoiceMem or implement a Memory Provider to replace storage and learning.