MemSlides Enables Precise Local Edits in AI‑Generated PPTs

Problem Motivation

AI‑generated slide decks often exhibit "whole‑deck drift" when users request local edits, causing instability in multi‑turn authoring. Users have persistent preferences (e.g., content density, visual style) and temporary constraints (e.g., "make later titles blue"). Without a dedicated memory mechanism these signals are either repeatedly injected into prompts or lost during revisions.

MemSlides Overview

MemSlides introduces a memory‑driven Slides Agent that treats personalized slide authoring as a stateful, multi‑turn problem. The system first builds an initial deck from a user‑profile memory, source material, and a selected template. Subsequent feedback updates the deck while preserving relevant memories.

Hierarchical Memory Framework

Long‑term profile memory : Stores stable cross‑task preferences (theme, content, visual, layout, template). At task start the appropriate profile bucket is selected and routed to working memory.

Working (session) memory : Holds active temporary preferences, carry‑over instructions, resolved targets, coverage status, and snapshot‑rebinding hints for the current deck. Conflict resolution follows the priority: explicit current feedback > task template > long‑term profile.

Tool memory : Captures execution experience at two granularities— round‑scope task experience (lessons from a specific modify job) and operation‑scope tool‑chain fragments (reusable reasoning‑tool observation snippets). This memory guides the agent to avoid redundant tool calls and to apply more reliable edits.

Plan‑Act‑Guard Pipeline & Scoped Local Revision

For each modification request MemSlides runs a three‑stage pipeline:

Plan : Constructs an execution contract that specifies the minimal effective slide region, target slide path, active rule identifiers, selector hints, and coverage requirements.

Act : Selects the appropriate editing tool based on the contract and applies the minimal edit to the identified region.

Guard : Verifies that the patch aligns with the correct snapshot, that target coverage is satisfied, and that finalization is not premature.

This scoped approach limits the edit to the target region, preventing unintended drift in non‑target slides.

Experimental Evaluation

MemSlides was evaluated on two fronts:

Personalized generation : Using a multi‑persona, multi‑intent user‑profile bank, the system improved round‑0 persona alignment while maintaining competitive PPT quality.

Multi‑turn local editing : A diagnostic matched‑pair setting isolated the effect of tool memory. Injecting tool memory raised overall closed‑loop completion from 0.815 to 0.963, strict verification from 0.310 to 0.534, and reduced time to first correct edit from 609.5 s to 242.5 s. Pair‑level analysis showed heterogeneity across models and task difficulties, indicating that tool memory does not uniformly dominate.

Key Findings

User‑profile memory boosts persona alignment in the initial generation stage.

Working memory enables delayed preference carry‑over within a session.

Tool memory improves closed‑loop completion, verification strictness, and editing efficiency for local modifications.

Effective local editing requires both task success and minimal drift in non‑target regions.

Broader Implications

MemSlides reframes presentation authoring as a memory‑augmented interactive process. Persistent profile memory can encode sensitive user habits, and improper consolidation may retain stale preferences. Future agents should provide transparent mechanisms for memory inspection, editing, and deletion to protect privacy.

Resources

Paper: https://arxiv.org/abs/2606.17162

Project site: https://memslides.github.io/

Demo: https://memslides.com/

Code repository: https://github.com/huohua325/Memslides

HuggingFace entry: https://huggingface.co/papers/2606.17162