MemSlides: A Hierarchical Memory-Driven Agent Framework for Personalized Slide Generation with Multi-turn Local Revision
Personalized presentation agents with user profile memory, working memory, tool memory, and scoped slide-local revision.
Official website notice: memslides.com is the
official MemSlides trial website. memslides.org is
a fake website, is not operated by this project, and is not affiliated with
this repository.
⭐ If MemSlides is useful for your research or slide-generation workflow, please consider starring this repository to help others discover it.
Demo Video
https://github.com/user-attachments/assets/a92ab49e-bc5c-4e90-8c0a-0f23b08a8857
Overview
MemSlides treats presentation generation as a stateful authoring process rather
than a one-shot source-to-slides conversion task. It separates personalization
signals by lifetime: persistent user profile memory captures recurring
cross-job preferences, working memory carries active session constraints across
revision rounds, and tool memory stores reusable execution experience for
reliable localized editing.
Long-term memory stores intent-conditioned user profile memory for round-0
personalization and tool memory for reusable execution experience. Working
memory maintains active preferences, session state, and revision constraints
within the current deck. During revision, MemSlides projects user feedback onto
the smallest affected slide region and applies scoped local patches instead of
repeatedly regenerating the full deck.
Highlights
- Intent-conditioned user profile memory routes personalization by
presentation intent, then applies preferences over theme, visual style,
layout, template use, content strategy, and general presentation habits.
- Multi-turn working memory preserves temporary preferences, session
constraints, and edit-state records across feedback turns in the same deck.
- Tool memory retrieves prior task and tool-chain experience before similar
edit operations to reduce repeated execution failures.
- Scoped slide-local revision updates the smallest affected slide region
instead of repeatedly rewriting the full deck.
Evidence
- User profile memory supports persona-aware round-0 personalization by routing
intent-matched preferences into the current job.
- Working memory carries active session constraints and temporary preferences
across multi-turn revision.
- Tool memory stores reusable execution experience so future localized edits
can avoid repeated failures.
- Scoped local revision keeps the edit surface close to the requested element,
reducing unintended drift in already aligned slide content.
Quick Start
Install from source:
sudo apt-get update
sudo apt-get install -y libreoffice fontconfig fonts-noto-cjk poppler-utils
conda env create -f environment.yml
conda activate memslides
pip install -e ".[research]"
python -m playwright install chromium ffmpeg
python -m memslides.experiment --help
Run the built-in smoke suite:
python -m memslides.experiment run smoke_minimal \
--output-base .memslides/experiments \
--parallel 1
The same experiment can run inside the Docker environment:
docker compose build
docker compose run --rm memslides python -m memslides.experiment run smoke_minimal \
--output-base /app/.cache/memslides/experiments \
--parallel 1
smoke_minimal is only a small verification suite. Users can pass any local
suite YAML path or packaged suite name to python -m memslides.experiment run.
Configuration
MemSlides needs user-provided model and service credentials for real generation
experiments. Keep credentials outside git and provide them through environment
variables, .env, or a private YAML file selected with MEMSLIDES_CONFIG_FILE
or --config.
The packaged public config is src/memslides/memslides.yaml; its placeholders
are expanded from the current process environment when the YAML is loaded.
Generated outputs, caches, private YAML files, and credentials must not be
committed.
Repository Structure
src/memslides/ Runtime package and public Python entry points
src/memslides/experiment/ Paper reproduction experiments and reports
src/memslides/tools/ MCP tool services and startup wrappers
src/memslides/memory/ Working-memory, profile, retrieval, and storage code
tests/ Automated regression and contract tests
.memslides/ Ignored runtime workspaces and generated outputs
~/.cache/memslides/ User-level caches, logs, and shared memory state
The public command-line entry points are generate, revise, and
template induct. The experiment runner remains available through
python -m memslides.experiment.
The generated tool snapshot and two unused internal modules were removed from
the source tree. Use the MCP configuration and public CLI instead of importing
memslides/.tools.json, memslides.memory.core.protocols, or
memslides.tools.ask_template_selection.
Development Checks
Install the development and research extras, then run the same checks used by
continuous integration:
pip install -e ".[dev,research]"
python -m compileall -q src tests
python scripts/check_source_hygiene.py
pytest -q
ruff check src tests
npm ci
npm run check:pptx-export
For Docker runs with a private YAML file:
docker compose -f docker-compose.yml -f docker-compose.private.yml run --rm memslides \
python -m memslides.experiment run smoke_minimal \
--output-base /app/.cache/memslides/experiments \
--parallel 1
The override maps ./memslides.private.yaml to
/run/secrets/memslides.private.yaml and sets
MEMSLIDES_CONFIG_FILE=/run/secrets/memslides.private.yaml inside the
container.
Experiment CLI
The suite runner is the main public entry point:
python -m memslides.experiment run smoke_minimal --output-base .memslides/experiments --parallel 1
python -m memslides.experiment report .memslides/experiments/smoke_minimal
python -m memslides.experiment personas
Core generation, revision, and template induction commands remain available for
scripted local use:
python -m memslides generate --instruction "Create a one-slide project summary" --num-pages 1
python -m memslides revise --workspace .memslides/session --feedback "Tighten the title"
python -m memslides template induct --template-file template.pptx
Security And Privacy
- Keep API keys in environment variables,
.env, or private YAML files.
- Do not commit
.env, .memslides/, generated workspaces, or private config
files.
- Network acquisition is optional and depends on user-provided search or model
credentials.
- External URLs and downloaded assets should be reviewed before presenting.
Citation
If you find MemSlides useful, please cite our paper.
@misc{jin2026memslides,
title={MemSlides: A Hierarchical Memory Driven Agent Framework for Personalized Slide Generation with Multi-turn Local Revision},
author={Ye Jin and Yangyang Xu and Jun Zhu and Yibo Yang},
year={2026},
eprint={2606.17162},
archivePrefix={arXiv},
primaryClass={cs.CL},
doi={10.48550/arXiv.2606.17162},
url={https://arxiv.org/abs/2606.17162},
}
License
See LICENSE and THIRD_PARTY_NOTICES.md.
MemSlides 是一个分层记忆驱动的智能体框架，把幻灯片生成当作一个有状态的创作过程来处理。它利用分层记忆（用户档案记忆、工作记忆、工具记忆）来个性化演示文稿，并采用局部范围的修订，而不是重新生成整个演示文稿。