Researchers Ye Jin, Yangyang Xu, Jun Zhu, and Yibo Yang introduced MemSlides, a hierarchical memory-driven agent framework designed to improve personalized presentation generation and multi-turn local revision. Unlike standard systems that append full history to a prompt, MemSlides organizes memory by temporal scope—long-term versus working memory—and functional role, specifically dividing long-term memory into user profile memory and tool memory.

User profile memory stores intent-conditioned profiles to enable round-0 personalization, while working memory retains active user preferences and session constraints across revision rounds. Additionally, tool memory stores reusable execution experiences to support reliable localized editing. This design enables the system to map user requests to the smallest affected slide region, preventing unintended changes in other parts of the presentation deck.

In controlled experiments, the framework improved persona-alignment judgments across a multi-persona and multi-intent profile bank. Diagnostic matched-pair testing showed that tool-memory injection enhances closed-loop modification behavior. The authors released the project code, a demo website, and project documentation on June 15, 2026, to support personalized slide generation that reliably incorporates user preferences and session-level constraints.

Researchers Ye Jin, Yangyang Xu, Jun Zhu, and Yibo Yang introduced MemSlides, a hierarchical memory-driven agent framework designed to improve personalized presentation generation and multi-turn local revision. Unlike standard systems that append full history to a prompt, MemSlides organizes memory by temporal scope—long-term versus working memory—and functional role, specifically dividing long-term memory into user profile memory and tool memory.

User profile memory stores intent-conditioned profiles to enable round-0 personalization, while working memory retains active user preferences and session constraints across revision rounds. Additionally, tool memory stores reusable execution experiences to support reliable localized editing. This design enables the system to map user requests to the smallest affected slide region, preventing unintended changes in other parts of the presentation deck.

In controlled experiments, the framework improved persona-alignment judgments across a multi-persona and multi-intent profile bank. Diagnostic matched-pair testing showed that tool-memory injection enhances closed-loop modification behavior. The authors released the project code, a demo website, and project documentation on June 15, 2026, to support personalized slide generation that reliably incorporates user preferences and session-level constraints.