Long-Term Memory

Long-term memory helps the assistant remember you across conversations. The default implementation uses Markdown files without a separate database.

Remember, Inspect, and Delete

Try:

• “Remember that I now live in Hangzhou.”
• “What have you remembered about me?”
• “Change my home city to Suzhou.”
• “Forget that address.”

Preferences are stored in <data-dir>/USER.md; facts and decisions in <data-dir>/MEMORY.md. A new conversation does not clear these files. Conversation edits apply immediately; direct file edits apply in the next voice session.

Lists and reference documents have separate features: Lists and Reminders, Knowledge Library.

Automatic Extraction

After a session ends, the default Markdown provider can use a text model to save missing long-term information: explicit interaction preferences in USER.md and stable facts or decisions in MEMORY.md. It cannot change ASSISTANT.md.

The default is DashScope qwen-flash using DASHSCOPE_API_KEY. Without an available key, automatic extraction is disabled; explicit memory-tool requests remain available. Disable extraction or configure another OpenAI-compatible text service:

QWEN_AUDIO_MEMORY_AUTO=off

Setting	Purpose
QWEN_AUDIO_MEMORY_MODEL	Text model for extraction
QWEN_AUDIO_MEMORY_BASE_URL	OpenAI-compatible service endpoint
QWEN_AUDIO_MEMORY_API_KEY	Credentials for that service

Extraction incurs additional model usage. It learns only from new conversation, not replayed history. After a successful memory edit through a tool or client, stale learning results cannot restore the old content. Extraction can still make mistakes; review and correct memories periodically.

Recall Earlier Conversations

Disabled by default. Enable topic summaries after sessions; the default retention is 90 days:

QWEN_AUDIO_SESSION_DIGEST=on

Ask “What was the project we discussed a few days ago?” Summaries are retrieved on demand; they are not recordings or verbatim transcripts.

A summary may record earlier work, but does not freeze its status. Available status comes from the current task ledger. Once a record expires, recall can describe what was discussed or requested, not verify current progress.

VoiceMem and Custom Providers

Optional VoiceMem takes over memory, retrieval, and session learning. Install it separately; the core npm package does not include it:

QWEN_AUDIO_MEMORY_PROVIDER=voicemem
VOICEMEM_PYTHON=/absolute/path/to/python
VOICEMEM_SIDECAR=/absolute/path/to/voicemem-sidecar.py
VOICEMEM_INPUT_MODE=text

text reuses transcription; audio sends turn-scoped user audio to VoiceMem. Switching providers does not migrate or delete the other provider's data. Back up first.

See Memory Provider for developer interfaces.

Privacy and Diagnostics

Relevant memories become voice-model context. Automatic extraction also sends conversation to the configured text service. Do not store passwords, API keys, verification codes, or tokens.

<state-dir>/memory-audit.jsonl contains learning diagnostics. Review logs before sharing. See Local Logs for standard log locations and rotation.