#1 HF PAPERS THIS WEEK · 171 UPVOTES

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

The Problem: Current AI presentation generators struggle with consistency and targeted edits. If you ask an AI to create a slide deck, it might do a decent job on the first try. However, when you ask it to revise something - like changing a specific data point or tweaking the tone on slide 3 - it often forgets your initial brand guidelines, loses context, or frustratingly regenerates the entire deck from scratch, wiping out previous manual adjustments.

The Breakthrough: MemSlides solves this by giving AI a "hierarchical memory" system, much like a human designer's workflow. It divides the AI's memory into three distinct functions:

  • User Profile Memory (Long-term): Stores persistent brand guidelines, formatting preferences, and user personas so the AI gets it right on "round zero."
  • Working Memory (Short-term): Tracks the ongoing conversation, active preferences, and new rules established during the current editing session.
  • Tool Memory: Acts as a library of reusable past experiences, teaching the AI exactly how to reliably execute specific technical formatting commands.

Crucially, MemSlides pairs this memory structure with scoped local revision. When a user requests a change, the AI pinpoints and updates only the smallest affected area instead of rewriting the entire presentation.

Why This Matters: This moves AI from being a one-shot "text-to-slide" novelty into a reliable co-authoring partner. By separating user identity from session context and mechanical editing skills, the AI becomes significantly more consistent. It aligns perfectly with complex brand personas, remembers instructions over a long back-and-forth chat, and makes precise, surgical edits without breaking the rest of the document.

Business Impact: For executives and product builders, this framework provides a blueprint for the next generation of enterprise productivity tools. The core concepts here extend far beyond presentation slides - they are highly applicable to any complex document generation, like drafting marketing collateral, legal contracts, or financial reports. By locally editing only the parts of a document that need fixing, companies can drastically reduce AI compute costs (inference overhead) while finally delivering the deeply personalized, multi-turn AI assistants that enterprise knowledge workers actually need.

Generated by Gemini