Reproduce “What to Remember, What to Reveal: Privacy-Aware Memory for Conversational Agents” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Wenjie Wang, Wenhe Si, Xinyue Xu, Yue Xu
Researchers at ShanghaiTech University developed SP-Mem, a privacy-aware memory architecture for conversational agents that isolates sensitive user data while maintaining personalization capabilities, alongside a benchmark to evaluate both utility and privacy adherence. This system effectively reduces unnecessary exposure of private information to rates as low as 0.33% in preference-only tasks, compared to 16.00% in full-context baselines, while preserving response quality.
Key insights
- Effective personalization in conversational agents can be achieved by decoupling memory utility from the exposure of exact private values.
- A hybrid memory architecture combining vector and graph representations provides complementary benefits for both general task completion and personalized understanding.
- Comprehensive privacy management throughout the entire lifecycle of sensitive data in long-term memory is essential, moving beyond single-stage protection.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.