Ye Jin, Yangyang Xu, Jun Zhu, Yibo Yang
4 min
Abstract
Personalized presentation generation requires more than conditioning on a current prompt or template: agents must preserve stable user preferences across tasks, retain newly introduced preferences and constraints during multi-turn revision, and carry out local edits reliably. We propose MemSlides, a hierarchical memory framework for personalized presentation agents that separates long-term memory from working memory and further divides long-term memory into user profile memory and tool memory. User profile memory stores intent-conditioned profiles for round-0 personalization, working memory carries active preferences and session constraints across revision rounds, and tool memory stores reusable execution experience for reliable localized editing. MemSlides pairs this memory design with scoped slide-local revision, so targeted updates act on the smallest affected region instead of repeatedly regenerating the full deck. In controlled experiments, user profile memory improves persona-alignment judgments on a multi-persona, multi-intent profile bank, tool-memory injection improves closed-loop modify behavior in diagnostic matched-pair settings, and qualitative cases illustrate working memory's ability to carryover preferences. Taken together, these results suggest that effective personalization in presentation authoring depends on separating persistent user profiles, session-level working memory, and reusable execution experience across generation and localized revision.
Sam: How does the system actually decide what counts as "the smallest region"? Is there a structured process for that?
Alex: There is. They use what they call a "Plan-Act-Guard" pipeline — three steps that happen in sequence for every edit. First, the system draws up something like a contract: a precise description of what needs to change and what must not be touched. Then it carries out the edit. Then — and this is the part that makes it reliable — a separate checking step verifies that the change was applied correctly and that nothing else was accidentally altered.
Sam: So the "Plan" sets the boundaries, the "Act" makes the change, and the "Guard" is the quality check that catches anything that slipped through.
Alex: Precisely. And the reason that three-step structure matters is that it separates intention from execution. The system commits to a plan before it acts, which makes it much harder for the edit to quietly expand beyond its original scope. It's a bit like a surgeon marking exactly where to cut before picking up the scalpel — the discipline happens before the action, not during it.
Sam: It sounds like the whole framework is less about raw capability and more about discipline — giving the AI a clear sense of what to pay attention to and what to leave alone.
Alex: That's the central argument of the paper, yes. The researchers suggest that the reliability problem with current AI presentation tools isn't primarily about the AI being insufficiently powerful. It's about the AI lacking the organisational structure to manage context well. A persistent memory of your preferences, combined with a disciplined approach to making only the changes that are needed — that combination is what makes the tool feel genuinely useful over time, rather than impressive once and frustrating thereafter.
Sam: So it reframes slide editing as an ongoing, iterative conversation rather than a one-shot task you have to restart every time.
Alex: That's a good summary. And it points to a broader shift in how we might think about AI assistants — not as tools that generate things on demand, but as systems that build up a working understanding of you and your work over time. Thanks for listening to ResearchPod.