A new arXiv paper introduces MemSlides, a hierarchical memory framework for personalized presentation-generation agents. The system separates long-term memory, working memory, user-profile memory, and tool memory.
That structure is designed to preserve stable preferences across tasks while also carrying new constraints through multi-turn revisions. Presentation agents need that continuity because users often make local edits while expecting the deck style and intent to stay consistent.
The paper points to a broader agent-design pattern: personalization requires memory that is structured by purpose, not just a single accumulated context buffer.