MemSlides Agent Tops Hugging Face Leaderboard: A Powerful AI for Personalized PPT Creation

MemSlides: An AI Slides Agent with Memory‑Driven Generation

MemSlides is a Slides Agent framework designed to address the pain points of AI‑generated PPTs. It creates personalized slide decks by remembering users' stable preferences and by handling multi‑round, localized edits without losing context.

Memory Architecture

The system separates memory into two hierarchical layers:

Long‑term memory stores cross‑task, stable user preferences such as low text density, heavy use of diagrams, or a habit of listing to‑do items on the final slide.

Working memory records temporary constraints, current task goals, and the status of ongoing edits, ensuring that instructions like “use blue titles for all new pages” persist across rounds until fulfilled.

In addition, tool memory captures fine‑grained operation experience (single‑round outcomes and step‑level actions) so that the agent can avoid repeating past mistakes during similar edits.

Scoped Slide‑Local Revision

To prevent over‑editing, MemSlides introduces a scoped revision mechanism that follows a three‑step “plan‑execute‑guard” workflow:

Plan : Translate the user’s request into a precise execution contract that defines which slides and which regions are to be modified.

Execute : Apply only the minimal, targeted changes dictated by the contract.

Guard : Verify that the modifications are complete, that no unintended areas were altered, and that the edit respects the established boundaries.

This process ensures that multi‑round revisions remain coherent and that already‑correct content is protected.

Experimental Evaluation

MemSlides was evaluated on three key metrics:

Closed‑loop completion rate (correct and finished edits) rose from 0.815 to 0.963.

Strict validation pass rate (meeting high‑quality standards) increased from 0.310 to 0.534.

First‑correct edit time dropped dramatically from 609.5 seconds to 242.5 seconds.

These results were obtained from diagnostic paired‑edit experiments that compared versions with and without tool memory. The data show that limiting the edit scope and reusing tool experience markedly improves both efficiency and reliability.

Conclusion and Outlook

MemSlides demonstrates that AI agents can move beyond one‑shot PPT generation toward sustained, multi‑round collaboration by correctly remembering long‑term preferences, preserving temporary constraints, and safeguarding edit boundaries. The framework’s memory‑layered design and scoped revision mechanism provide a blueprint for future AI‑driven document creation, code modification, data analysis, and enterprise knowledge‑management tasks.