MemFold research website

Website: https://MemFold.github.io/

Repository: https://github.com/MemFold/memfold.github.io

A responsive research blog introducing the paper, with model-selectable benchmark results, expandable figures, and arXiv, GitHub, and Hugging Face resource buttons. No build step, external fonts, analytics, or JavaScript dependencies are required.

• index.html: article, authors, links, and initial results table.
• styles.css: desktop and mobile layout.
• script.js: model selector, figure viewer, and reading navigation.
• assets/: paper figures, the project logo, and institutional logos.

Run python3 -m http.server 8000 --directory website from the manuscript root, or python3 -m http.server 8000 from the standalone website repository.

GitHub Pages publishes the main branch, root folder. Push this directory's contents to the website repository root; every push deploys automatically. .nojekyll keeps the site as plain static files.

Article text and results are verified against the published paper, arXiv:2609.36435v1 (September 29, 2026). Website figure numbers match the paper. Benchmark values come from arxiv/main_table.tex; ablations from arxiv/5_ablation.tex; the case and memory analysis from arxiv/5_discussion.tex. Experimental plots are redrawn as page-styled SVG using the unchanged data in tools/figure-data.json. tools/render_figures.py generates desktop and mobile training plots plus memory analyses (requires matplotlib). Colors match the page, text remains vector text, and mobile layouts stack the training panels. The framework diagram retains its PNG rendering, as requested. The logo uses SVG paths and gradients; figure zooms use the same SVG files as the article. The full manuscript PDF is not included in the public website.

Resource URLs match the links embedded in the published PDF:

• Paper: https://arxiv.org/abs/2609.36435
• Code: https://github.com/Johnny221B/memfold
• Model: https://huggingface.co/Johnny221B/memfold

The citation includes the arXiv identifier, subject class, and DOI.

Figures fill the content area to the right of the desktop reading outline. All figure containers have transparent backgrounds and no borders. The framework PNG has a transparent alpha channel. Its two full-page white vector background paths were removed from the source PDF export; the remaining native drawing instructions are preserved in tools/framework-transparent.svg. The two memory-analysis plots and budget cost plot share a row and identical 367.2 × 298.8 SVG viewBoxes. On smaller screens they stack vertically.

The Discussion figures include training efficiency, budget accuracy, budget cost, memory reliance, and the personalization example. tools/render_case.py extracts all three responses from arxiv/5_discussion.tex and lays out responsive SVGs using website typography. It verifies each paragraph against the extracted source. The standalone snapshot is tools/case-text.json. Keep every word, punctuation mark, label, and answer token count unchanged when restyling.

The site uses a near-white background, neutral text and borders, and blue accents. Figure enlargement is available by clicking the images, without additional visible interface labels.

The desktop cover fills the first viewport. The article begins below the cover divider. Institutional logos follow the supplied TAPS-DLM website layout; logo provenance is recorded in assets/logos/README.md.

The cover title, article copy, lists, section headings, and subheadings all use Georgia / Times New Roman / serif. Captions, table metadata, and generated SVG labels retain Arial / Helvetica / sans-serif.

Plot typography is defined once in the TYPE mapping in render_figures.py: 14 pt panel titles, 12 pt axis labels, 11 pt ticks / annotations / legends, and 10 pt legend group labels. SVG sizes scale with their responsive canvases. The case figure uses 18 px response text on desktop and 17 px on mobile. The framework retains the original artwork's typography.