Source-linked AI summary
MemSlides: A Hierarchical Memory Driven Agent Framework for Personalized Slide Generation with Multi-turn Local Revision
Ye Jin, Yangyang Xu, Jun Zhu, Yibo Yang
TL;DR
Personalized slide agents still struggle to preserve user preferences across tasks and revisions while making reliable local edits. MemSlides separates profile, working, and tool memory and applies scoped slide-local revision; controlled evaluations show improved persona alignment and local-modification reliability.
Problem
Presentation agents lack persistent personalization and robust multi-turn local revision, while full-deck regeneration can overwrite aligned content and increase context pressure.
Method
MemSlides separates user profile, working, and tool memory while routing active session state to scoped slide-level edits instead of full-deck regeneration.
Results
MemSlides improves persona alignment and local-modification reliability; tool-memory injection raises Closed-Loop Completion from 0.815 to 0.963.
Takeaways & Limitations
Effective personalized presentation authoring depends on separating persistent profiles, session working memory, and reusable execution experience across generation and localized revision.
Takeaways & Limitations
Evidence is limited to controlled persona judgments, diagnostic matched-pair modify settings, and qualitative cases rather than real-user deployment studies.
Abstract
from arXiv · showhide
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.
1 Introduction
MemSlides addresses fragile personalized presentation revision by combining scoped slide-local editing with hierarchical memory. Its long-term and working memories preserve persistent user preferences, reusable execution experience, and active session constraints across generation and revision.
- Motivation: Existing presentation agents can generate polished decks, but they do not explicitly model user-specific personalization and still make high-quality presentation creation time-consuming and cognitively demanding.Personalization also requires users’ long-term preferences in organization, styling, and revision to persist across different intents.
- Motivation: Multi-turn personalization is fragile because preferences often emerge during revision, while existing agents re-contextualize or regenerate large deck portions for small edits.These edits compete with deck state and feedback history for limited context, weakening local modification.
- Method: MemSlides projects each feedback request onto the smallest affected slide region, reads a structured local snapshot, and performs targeted updates instead of repeatedly rewriting the full deck.The bounded repair surface includes local layout structure, available selectors, and exposed style rules.
- Method: MemSlides separates long-term memory into intent-aware user profiles and reusable tool experience, while working memory carries session state and temporary feedback across revision rounds.User profile preferences span theme, content, visual, layout, template, and general dimensions, then route into working memory when a job starts.
- Evaluation: Experiments assess deck quality, instruction satisfaction, preference alignment, and persona-alignment judgments, with qualitative comparisons illustrating preference alignment.The paper evaluates personalization alignment through newly developed persona-alignment metrics.
2 Related Work
Related work spans slide generation, memory- and tool-using agents, and personalized generation, establishing foundations for adaptive authoring and preference-aware systems. These lines of research motivate combining controllable presentation generation with persistent context and user-history-aware personalization.
- Slide generation: Slide-generation research has progressed from document compression and structured summarization toward audience adaptation, editability, task-time preference inference, and visual refinement.Presentation authoring also incorporates controllable layout and design generation through code-like representations, in-context prompting, layered or diffusion-based modeling, and visual preference modeling.
- Memory and tool-using agents: Memory- and tool-using agent research shows how stored context, structured memory management, interleaved reasoning and actions, modular tools, and feedback can support generation and execution.Prior studies address persistent memory, reflection, structured updates, long-term/short-term memory management, API use, coordination, and learning from feedback.
- Personalized generation and evaluation: Personalized generation has advanced from explicit persona conditioning to profile- and history-aware generation, with recent work framing personalization as agentic, retrieval-aware, and preference-sensitive.In visual domains, personalized visualization recommendation and DesignPref demonstrate that persistent expressive or design preferences can be learned from user history.
3 MemSlides
MemSlides formulates personalized presentation generation as stateful, multi-turn authoring with separate personalization signals and scoped revision. Its hierarchy combines persistent user profiles, session working memory, and reusable tool-execution experience to preserve preferences while editing only affected slide regions.
- Stateful authoring: MemSlides models authoring as round-0 generation followed by feedback-driven state updates and deck edits across revision rounds.The session state stores active constraints and edit intentions, while the objective includes preserving slides that are already aligned.
- Memory hierarchy: The framework separates recurring user preferences, task-local template constraints, and temporary session requirements by lifetime and scope.User profile memory covers themes, visual styles, layouts, density, and design conventions; templates constrain the current deck; session state carries turn-specific requirements.
- Scoped revision: Localized revision avoids repeated full-deck regeneration because requests may target one slide, multiple slides, or deck structure, reducing overwriting and unnecessary drift.The execution protocol converts each request into a contract specifying scope, target slide paths, active rules, selector hints, and coverage requirements.
- Working memory: Working memory carries temporary preferences, carryover instructions, edit-state records, coverage status, and snapshot-rebinding hints across multi-turn Plan–Act–Guard execution.It also buffers round-level tool-memory signals before transferable experiences are consolidated into long-term tool memory.
- User profile memory: User profile memory routes intent-matched, request-compatible preferences into active memory, supersedes explicit conflicts, and consolidates only stable signals at job end.This lifecycle prevents transient requests from becoming persistent preferences while retaining transferable signals for future personalization.
- Tool memory: Tool memory records scope- and verification-linked execution experience to reduce repeated ineffective trials and known tool misuse during personalized editing.Its flow includes round-scope lessons, tool-error summaries, extracted patterns, and operation-scope reasoning–tool–observation fragments retrieved before similar future calls.
4 Experiments
Experiments show that MemSlides improves round-0 persona alignment while preserving competitive presentation quality, and that tool-memory injection strengthens closed-loop localized revision. Qualitative evidence further indicates that localized patches preserve non-target content and that repeated feedback can become reusable profile preferences.
- User profile memory: MemSlides improves persona alignment across model families, winning all dimensions against both baselines for GLM-5 and Gemini 3.1 Pro.With GPT-5, it remains ahead on several dimensions but does not uniformly dominate every dimension.
- General deck quality: Persona-alignment gains remain compatible with ordinary presentation quality, with OURS achieving the best GPT-5 average and remaining competitive on GLM-5.Gemini 3.1 Pro leads Style and Diversity but has lower Constraint, so the results do not show uniform dominance on every quality metric.
- Localized revision: Localized patch revision targets the requested element while preserving already aligned non-target slide content, unlike broader whole-slide rewriting.Figure 5 provides qualitative evidence for this locality distinction, while the modify evaluation pairs locality with completion, verification, and time constraints.
- Memory carryover: Across repeated jobs, local feedback cues become reusable profile preferences that later reappear as default slide-organization patterns.This is qualitative cross-job evidence rather than a separate quantitative metric; complementary within-session delayed carryover is reported in Appendix Figure 9.
- Tool-memory ablation: 0.963 versus 0.815 Closed-Loop Completion and 0.534 versus 0.310 Strict Verify show stronger closed-loop reliability with tool-memory injection.These results come from nine diagnostic matched modify pairs, where only tool-memory injection changes.
- Tool-memory ablation: 242.5s versus 609.5s Time to First Correct Edit and a 0.327× Core Tool Time Ratio indicate faster correct editing and less non-inspection tool work.The corresponding pair-level outcomes are reported as W-L-T-NA counts of 6-2-0-1 and 8-1-0-0.
5 Conclusion
MemSlides is a hierarchical memory framework for personalized presentation generation that separates user profile, active temporary, and tool memory. This design supports round-0 persona alignment and multi-turn localized revision, with experiments showing gains in persona alignment and local modify reliability.
- MemSlides introduces a hierarchical memory framework for personalized presentation generation.Its memory design separates user profile memory, active temporary memory, and tool memory.
- Separating these memory types supports round-0 persona alignment and multi-turn localized revision.
- Controlled experiments show improved persona alignment and diagnostic gains in local modify reliability.
6 Limitations
The evidence is limited to controlled persona-alignment judgments, diagnostic matched-pair modify settings, and qualitative working-memory cases using proxy profile banks and edit requests. Broader human studies, randomized edits, and stronger memory safeguards remain future needs.
- Evidence scope: Evidence is scoped to controlled persona-alignment judgments, diagnostic matched-pair modify settings, and qualitative working-memory cases.These settings do not establish performance in broader deployment contexts.
- External validity: The profile bank and edit requests are proxies rather than real-user deployment studies.This limits direct evidence about real-user behavior and needs.
- Future work: Future work should add broader human studies, randomized edit sets, and stronger memory consent, deletion, and sensitive-preference safeguards.The proposed safeguards address privacy and control over stored preferences.
A Appendix
The appendix presents protocol details, supplemental quantitative tables, and qualitative examples explaining how memory signals appear in generated presentations. Its qualitative figures are diagnostic, while aggregate claims remain grounded in quantitative tables.
- The appendix provides protocol details for the study.
- It includes supplemental quantitative tables supporting the paper’s aggregate claims.
- Qualitative examples illustrate memory signals in generated presentations and serve as diagnostic figures.
A.1 Evaluation Protocol Details
The evaluation protocol isolates memory effects by controlling prompt leakage, arm ordering, source-material matching, and execution-trace measurement. Profile memory is judged through blinded, rubric-based round-0 deck comparisons, while tool memory is evaluated through matched localized-edit executions and defined process metrics.
- Protocol controls: Protocol controls target prompt leakage, arm ordering, and unmatched source materials so reported metrics remain attributable to the corresponding memory condition.These controls are documented as part of the evaluation protocol for Section 4.
- Profile-memory evaluation: Profile-memory judging uses page-aligned round-0 deck images with anonymous arm labels, exposing judges only to the target persona summary, decks, and dimension rubric.The original prompt, parsed intent, system identity, and memory condition are hidden.
- Profile-memory evaluation: Persona-alignment judgments score content, structure, visual, and another rubric dimension on four 0–10 dimensions, with structure excluding template matching.Content covers selection, evidence type, emphasis, and wording; structure covers page order and page-type/layout fit.
- Tool-memory evaluation: Tool-memory evaluation fixes the source deck, model family, persona, and modify request within each pair, varying only whether tool memory is injected.Trace-derived metrics include edit completion, verification after change, finalization, time to first correct edit, and core tool time.
- Tool-memory evaluation: Table 3 defines Closed-Loop Completion and Strict Verify After Change as higher-is-better metrics, alongside two lower-is-better efficiency metrics.Closed-Loop Completion requires successful local editing, verification, and finalization; Strict Verify After Change rewards prompt local edit-and-check behavior.
A.2 Baseline Conditions and Prompt/Profile Separation
The evaluation separates persistent profile memory from prompt-level task conditioning by comparing matched conditions that differ in memory access. Judging also withholds prompt and template-related information, while Structure excludes template matching to assess persona-conditioned organization and layout fit.
- Controlled comparison: Matched memory-injected and control conditions use the same source material, model family, target persona, task prompt, and generation pipeline.The control withholds long-term profile memory, session preference memory, and reusable tool experience from the generation context.
- Controlled comparison: Profile memory is evaluated as an additional persistent signal because the task prompt already contains the source-specific request and role intent.It represents structured preferences accumulated across controlled authoring jobs rather than replacing the user request.
- Judging protocol: The persona-alignment judge receives the target persona summary and rendered deck images but not the original prompt, parsed task intent, system identity, or memory condition.This protocol limits prompt or template leakage from dominating the score.
- Judging protocol: The Structure metric excludes template matching, so the main profile table evaluates persona-conditioned organization and layout fit rather than template retrieval accuracy.SlideTailor is included as an external reference/template-conditioned personalization baseline under the same judging protocol.
A.3 Compute Resources and Runtime Accounting
Reported experiments were inference-time agent evaluations orchestrated on a local Linux workstation, with runtime accounting tracking provider usage, local tool outcomes, active time, and wall-clock span. In the diagnostic tool-memory setting, memory injection used less core tool time and fewer core tool calls than no injection.
- Compute resources: All experiments evaluated inference-time agents without model training or fine-tuning, using a local Linux workstation for orchestration and slide-processing tasks.The workstation has 48 Intel Xeon Silver 4214R CPU threads and 503 GiB of system memory; experiments were primarily API-bound rather than GPU-bound.
- Runtime accounting: Runtime accounting records model-response usage, local tool-call outcomes, tool active time, and first-to-last-event wall-clock span.Provider-reported prompt and completion tokens are used for reproducibility accounting, not as a normalized price estimate.
- Diagnostic tool-memory setting: 110.5 seconds versus 354.8 seconds: memory-injected runs used less core tool time than no-injection runs in the diagnostic matched-pair setting.Inspection and markdown-conversion tools were excluded from these process metrics.
- Diagnostic tool-memory setting: 779 versus 878: memory-injected runs contained fewer core tool calls than no-injection runs under the same diagnostic matched-pair exclusion.The comparison excludes inspection and markdown-conversion tools.
A.4 Profile-Bank Construction
The profile bank is a controlled 30-entry testbed spanning ten occupation-style personas and three role-intent buckets, constructed through authoring interactions and seeded completion. Seeded completion fills sparse structured fields with provenance while adding no synthetic interaction, tool, or template-use records.
- Profile-bank scope: The bank contains 30 persona-intent entries formed from ten occupation-style personas and three role-intent buckets.Personas include postsecondary teacher, software developer, management analyst, marketing manager, graphic designer, training and development specialist, financial manager, operations manager, medical and health services manager, and legislator.
- Construction stages: Each entry first receives controlled authoring interactions on the same source material under different role-intent prompts, producing initial profile evidence.This first construction stage is designed to create profile evidence specific to each persona-intent combination.
- Construction stages: A seeded completion step fills missing structured fields using stable persona prompts, an occupation-grounded role-preference registry, existing profile signals, and suite intent definitions.The completion step is applied because some accumulated entries remain sparse in structured fields.
- Construction stages: Seeded completion updates only structured profile entries and does not create synthetic interaction episodes, tool experiences, or template-usage records.Each filled field receives provenance tags identifying whether it came from the seed prompt, role-preference registry, current profile signal, or suite intent definition.
A.5 DeepPresenter-Style Quality Evaluation … A.11 Broader Impacts and Responsible Use
Across quality, personalization, template, memory, and responsible-use analyses, the paper evaluates whether memory-driven personalization preserves presentation quality while improving alignment and localized editing. It also documents practical benefits, asset constraints, and risks from persistent personalization and misleading content generation.
- A.5 DeepPresenter-Style Quality Evaluation: DeepPresenter-style evaluation checks whether profile-memory injection preserves ordinary presentation quality alongside persona-alignment gains.Content and Style use released PPTEval-style prompts, Constraint is reimplemented for this task suite, and Diversity uses DINOv2 embeddings with normalized Vendi scores.
- A.5 DeepPresenter-Style Quality Evaluation: Across three model families, OURS remains competitive on general-quality metrics, with Content above SlideTailor and close to or above DeepPresenter without injection.The passage identifies Content as the most stable pattern across model families.
- A.6 Additional Persona-Alignment Judgment Results: Profile-memory injection improves GPT-5 persona alignment by 2.42 Overall points, including gains of 3.30 Content, 2.30 Structure, 3.17 Visual, and 2.43 Specificity.Table 7 reports these ten-persona results on a 0–10 judge scale, with Overall averaging Content, Structure, and Visual.
- A.7 Template-Guided Generation Examples: Template-guided examples show templates imposing layout, palette, typography, and visual-organization constraints while persona conditioning adapts slide content and emphasis.Each example pairs a selected template slide with a generated slide for the same source paper.
- A.8 Tool-Memory Pair-Level Details: Tool-memory matched-pair results show that most pairs improve or tie on closed-loop completion across the diagnostic setting rather than only for one model.Gemini 3.1 Pro uses a fixed graphic-designer hard-modify family, exposing difficult cases without favorable-outcome pair selection.
- A.8 Tool-Memory Pair-Level Details: The paired robustness check finds strongest evidence for Strict Verify and Core Tool Time Ratio, while Closed-Loop Completion and First Correct Edit remain directionally favorable.The exact one-sided sign test uses pair-level wins and losses after excluding ties and unavailable pairs, within the controlled diagnostic protocol.
- A.9 Additional Qualitative Memory Cases: Working-memory cases illustrate delayed preference carryover when a rule stated earlier becomes actionable only after a later edit.Memory injection retrieves the stored rule at the trigger turn, whereas no-memory behavior relies mainly on local context.
- A.10 Existing Assets and Use Conditions: Personalized presentation agents may reduce authoring effort and improve audience, role, style, and localized-revision consistency, but persistent profiles can preserve sensitive, outdated, or unintended preferences.The paper also warns that presentation generation may produce persuasive but misleading materials or over-adapt framing to target audiences, while experiments use existing systems, hosted APIs, visual models, and author-created templates without redistributing third-party assets.