MemSlides: A Hierarchical Memory Driven Agent Framework for Personalized Slide Generation with Multi-turn Local Revision

Researchers introduce MemSlides, a novel agent framework designed to enhance personalized presentation generation by implementing a hierarchical memory system that ensures preference stability and precise local revisions across multi-turn interactions.

Overcoming the Challenges of Personalized Presentation Generation

Generating high-quality, personalized presentations extends beyond simple prompt conditioning or template application. Current AI agents often struggle with maintaining consistent user preferences across different tasks and retaining specific constraints introduced during iterative, multi-turn revision processes. Furthermore, performing reliable local edits without disrupting the overall structure of a presentation remains a significant technical hurdle.

The MemSlides Architecture: Hierarchical Memory Management

To address these limitations, the authors propose MemSlides, a framework that utilizes a hierarchical memory structure to decouple different types of information. This architecture separates long-term memory from working memory to optimize how the agent processes both static user preferences and dynamic task constraints.

Long-Term Memory Segmentation

The long-term memory within MemSlides is further subdivided into two specialized components:

• User Profile Memory: Stores stable, overarching user preferences to ensure consistency across multiple presentation projects.
• Tool Memory: Manages the technical capabilities and operational constraints of the tools used for slide generation, ensuring the agent interacts with the generation engine efficiently.

Working Memory and Local Revision

By utilizing a dedicated working memory, MemSlides can track newly introduced preferences and specific constraints during multi-turn revisions. This allows the agent to perform precise local edits, ensuring that modifications to a specific slide or element do not negatively impact the rest of the document.

Note: Due to the limited nature of the provided source text, specific evaluation metrics, dataset details, and quantitative results of the MemSlides framework are not available.