Persistent User-Customizable Assistants (Agent Profiles) like Custom GPTs / Gemini Gems

[Muhammad313a]

Is your feature request related to a problem? Please describe.
Currently, users must restate system constraints, role, and reasoning style in every new session. This leads to inconsistent behavior across conversations, extra cognitive overhead, and weaker outcomes for long-term technical workflows (coding/reasoning/research).

Describe the solution you'd like
Add support for persistent user-customizable assistant profiles (agent profiles), similar in concept to Custom GPTs (ChatGPT) and Gems (Google Gemini).
A profile should allow users to define:

• persistent system-level instructions (behavior/policy),
• a stable role (e.g., strict tutor, code reviewer, formal analyst),
• optional attached reference documents / knowledge sources,
• consistent behavior across sessions.

Describe alternatives you've considered

• Using a saved/pinned prompt template: helps, but still requires manual reuse and does not reliably enforce consistent behavior across sessions.
• Copy/pasting system instructions each time: works but is repetitive and error-prone.

Additional context
This is not model fine-tuning. It is a persistence layer for assistant behavior and optional knowledge attachments. Given DeepSeek’s strengths in reasoning and code-related tasks, this capability would enable disciplined, repeatable workflows and can reduce hallucinations by keeping the assistant constrained and consistent.

[Muhammad313a]

Thanks! Just to clarify: this request is about the DeepSeek product/platform (chat UI / API experience) — persistent custom assistants (agent profiles) similar to Custom GPTs / Gemini Gems — not about changing the V3 model weights.

If this repo isn’t the best place for product/platform feature requests, could you please point me to the correct repository or feedback channel? I’m happy to repost it there.