Ever fix one slide and watch the agent regenerate the whole deck and break everything? This work solves that with hierarchical memory plus local revision 🗂️ Title: MemSlides: A Hierarchical Memory Driven Agent Framework for Personalized Slide Generation with Multi-turn Local Revision URL: ❓ What's new? 💡 It gives a slide-generation agent a hierarchical memory that cleanly separates persistent preferences (user profile), session-level working memory, and reusable tool experience. It also edits only the target slides instead of regenerating the full deck. ❓ How does it keep local edits from drifting? 💡 A Plan–Act–Guard loop. Plan turns a request into an explicit "execution contract" (target paths, active rules, coverage requirements); Act picks tools like batch CSS over shared selectors or patch operations; Guard binds patches to content hashes and blocks premature finalization until all targets are covered. ❓ Do one-off instructions leak into permanent preferences? 💡 No — at job end, intent-aware consolidation writes back only stable signals, preventing transient "just this round" requests from becoming persistent preferences. ❓ Does it work? 💡 In blind review, personalization beats DeepPresenter on every dimension (e.g. Visual +1.66). With tool memory, closed-loop completion hits 0.963 and time-to-first-correct-edit drops 609.5s → 242.5s (~60% faster). Designing agent memory by role rather than one flat store feels broadly useful for multi-turn editing tasks. #AIAgents# #LLM#