MemSlides: Hierarchical memory AI for personalized slide generation

MemSlides tackles a core challenge in AI-powered slide creation: keeping personal preferences consistent across tasks and revisions. The framework introduces a three-tier memory hierarchy. Long-term memory stores stable user profiles and reusable tool knowledge (e.g., how to execute specific edits reliably). Working memory captures session-specific constraints and newly introduced preferences that carry over across revision turns. This design allows the agent to personalize slides from the first generation (round-0) based on stored intents, then refine them locally without rebuilding the entire deck.

For revision, MemSlides uses scoped slide-local edits that target the smallest affected region rather than regenerating the full presentation. Controlled experiments demonstrated that user profile memory significantly improves persona-alignment judgments on a multi-intent profile bank, tool memory injection enhances closed-loop modification in matched-pair tests, and qualitative examples show working memory's ability to carry preferences across turns. The approach marks a step toward authoring assistants that truly remember users, not just their last prompt.