MemoAI: video to subtitles, translation and dubbed audio on the desktop

A repository that ships binaries rather than source

Makememo/MemoAI has 1,066 stars and 108 forks, points at memo.ac as its homepage, and has 103 open issues against a default branch called `main`. The last push landed on 2026-09-12 and the repository is not archived. What the project does not have is code. The tree contains exactly two entries, `.github/` and `README.md`, and neither the README nor the releases describe an open-source build.

That shape tells you what kind of project this is. Memo AI is a commercial desktop application that happens to use its GitHub repository as the download page, the changelog, and the support channel. The README makes the platform explicit: Memo supports macOS Silicon, macOS Intel and Windows, and installation happens from the releases rather than from a package manager or a build step. The `license` field on the repository is empty, so there is no grant for redistribution.

For a reader deciding whether to try it, the useful part of the repository is therefore the release history rather than the file listing. Three releases in the second half of September 2026 give a much sharper picture of where the product is going than a source tree with two files would.

Three ways to get media into the app

The README documents three intake paths, and they differ in how much setup they cost you. The first is a link: copy a YouTube link or a podcast link, paste it into the Memo input box, and click transcribe. There is no command to copy and no flag to pass.

The second path is a local file. MemoAI handles audio and video formats such as MP4, MP3, AAC and M4A directly, and the README is explicit that no conversion step is needed first. It attaches one condition that will save you an afternoon of debugging: the name of the file being converted must not contain special characters, or MemoAI will not recognise it. That is a naming constraint, not a format limitation.

The third path skips transcription entirely. If you already have a subtitle file and only want it translated, you can upload the file directly, and the supported formats are SRT and VTT. This is the path most worth knowing about if you already run a transcription tool and only lack translation.

Both link and file intake have been worked on recently. Version 1.8.5 fixed the inability to fetch and transcribe online videos, tracked as issues 424 and 420, and version 1.8.6 added drag-and-drop of files onto the player so that media can be associated with a subtitle-only transcription job.

Model selection and paragraph shaping

Transcription runs against a model you pick from a picker. The README's only tuning advice concerns paragraph shape: if you want AI-style paragraph output rather than one line per utterance, you adjust the maximum number of words in the paragraph, usually 300. That is a small control with a large effect on the readability of the finished file, because subtitle timing and paragraph boundaries pull against each other.

The provider landscape has been moving. Version 1.8.4 reorganised the article-conversion model picker so models are grouped by provider, and added the ability to remember whether the last auto-fix transcription run was on or off. Version 1.8.5 added the ability to fetch the latest models online, so the list is no longer frozen at build time. Version 1.8.6 fixed an Ollama model list API call error, which means a locally served model list is a supported path rather than an accident.

The named providers that surface in these notes are Ollama for local models, Gemini and SiliconFlow for hosted ones. None of them is configured at install time in the way you would configure an API key in a library. The app is the configuration surface, which is convenient for a single user and limiting for anyone who wants a scripted pipeline.

Two free translation paths and the paid ones behind them

Memo ships two built-in free translations, Google and Microsoft, and the README states plainly that they meet the needs of daily use. That is the single most useful sentence in the document for anyone evaluating cost, because it means the basic loop of transcribe then translate requires no subscription.

If those are not enough, three alternatives are named: Volcano Translation, DeepL and AI Translation. Selection happens per job rather than per account, and the release notes show this area being debugged rather than merely offered. Version 1.8.6 fixed Gemini and SiliconFlow being unable to invoke translation, which tells you those paths exist and had a failure worth shipping a patch for.

The interface offers two levels of retry. If a whole paragraph is unsatisfying, paragraph translation re-runs it. If one line is wrong, line translation re-runs just that line. Neither is automatic quality control, and neither compares candidate outputs, so a reader who does not speak the target language has no easy way to know whether a fluent-looking paragraph is a faithful one.

Speech synthesis and the export formats that matter

Speech synthesis is the feature that separates Memo from a plain transcription tool. The app supports a variety of synthesis methods, and a translated language can be dubbed over the original media rather than sitting beside it as text. Version 1.8.5 fixed the inability to export the video and the synthesized audio track, which tells you the export path existed and was broken rather than newly added.

Text export covers the common subtitle formats, SRT and VTT, and the README claims this eliminates manual adjustment. Alongside the subtitle formats there is synchronized export with Markdown and other tools, which is aimed at people who keep their notes in an Obsidian-style vault. Version 1.8.4 fixed truncated content length when exporting to Obsidian, and also fixed an include-speaker checkbox that was being ignored during subtitle export.

Export naming became more configurable in version 1.8.6, which added multi-language selection and custom naming rules on the detail page. Between the two releases, the export surface moved from a single fixed filename to something you can shape per language, which is what you need the moment a project has five subtitle tracks in it.

What three releases in two weeks reveal

The release history is short enough to read in full, which makes it unusually informative. Version 1.8.4, published 2026-08-29, added folder-based resource management and spent much of its fix list on batch behaviour: batch task concurrency issues across folders, batch exports not taking effect when tasks were re-run, and the display state when resuming a translation that had been stopped partway through. It also fixed subtitle timing damage caused by merging duplicate lines during subtitle optimisation.

Version 1.8.5, published 2026-09-08, was smaller and concentrated on the editor: in-place search and replace within selected subtitle segments, and a backspace hint to merge the current line with the previous segment, tracked as issue 165. Version 1.8.6, published 2026-09-12, went back to the player, adding auto-locate to the current playback or editing position when switching clips, real-time progress tracking when re-transcribing selected continuous text, and a fix for follow-scrolling that only worked during playback.

Read together, that is a project spending its effort on long-session editing ergonomics and on batch correctness rather than on adding providers. The 103 open issues against 1,066 stars is the other signal worth holding onto: a large user base is pushing on a codebase that is not visible to them.

Editorial conclusion

MemoAI is at its best when a single person has a stack of recorded talks or podcasts and wants finished subtitles in several languages without touching a subtitle editor. The interesting engineering is not visible in the repository because there is no repository code: the only public artefacts are the README, an issue tracker with 103 open items, and the release page that carries the installers. The release notes from August and September 2026 show the feature work moving toward batch handling, folder organisation and export naming rather than toward new model providers. Start with one local file, keep the filename free of special characters as the README requires, and only move to the paid translation providers once the free Google and Microsoft paths have shown you what the transcript quality actually needs.

Frequently asked questions

Official sources

1. Issues
2. Makememo/MemoAI on GitHub
3. Project website
4. README
5. Releases