How it works
Where words come from
A word enters a MemDecks collection through exactly four doors: a Phrase Builder lesson, the Translator, a ready-made deck, or an Anki import. What fills the card behind it comes from several independent sources — a stored dictionary first, machine translation as the baseline for anything new, a separate romanization step, separately generated audio, and separately generated lexical detail. Any of the optional parts can be missing without the card being broken.
The four word sources
| Source | Pace | Who chose the word |
|---|---|---|
| Phrase Builder lesson | Five words per lesson | MemDecks, as part of an authored phrase |
| Translator, typed | One at a time | You |
| Translator, Next word | One at a time | MemDecks, from a topic pack you follow |
| Ready-made or imported deck | A whole deck at once | MemDecks, you, or whoever made the Anki deck |
These four are the complete list. Nothing else adds cards to a collection, and MemDecks never adds words you did not ask for — with one stated exception: a new account is subscribed to the Starter topic pack, so the pump has something to offer. A subscription proposes words; it does not save them.
The MemDecks word catalogue
Behind the pump sits a catalogue of English lemmas with per-language translations, difficulty scores and lexical detail. The 21 topic packs draw on it: 1,392 pack entries covering 1,302 distinct English words as of 2 September 2026.
The catalogue is why the pump can hand you words in difficulty order rather than alphabetically, and why card HP can differ per language for the same English word.
Where translations come from
MemDecks looks in its own stored dictionary first. A word already in the catalogue for your study language is served from there, which is why most lookups are instant and identical for everyone.
A word that is not in the catalogue is translated by Google's neural machine translation, which is the baseline for all 16 study languages. That result is what becomes the card. Later, an optional lexical pass can add senses, examples and synonyms on top, but it never silently replaces the baseline translation.
Where transliteration comes from
Transliteration is produced separately from translation, and only for the nine study languages that do not use the Latin alphabet. Armenian and Greek use deterministic rules written specifically for those alphabets — the same word always transliterates the same way. The other non-Latin languages use a general romanization service, with a fallback chain behind it.
Where audio comes from
Audio is generated once by a text-to-speech service and stored as a file, then replayed from storage. Armenian and Greek use a Gemini voice profile chosen for those languages; the other fourteen use the standard Google voice. Generation happens in the background, so a brand-new word can be silent for a few seconds after saving.
Cards imported from an Anki package may instead carry the audio that was in the original deck.
Where images come from
Cards created through the Translator get a photograph found on Unsplash for the English word. Ready-made MemDecks decks ship their own artwork instead, so their cards look consistent. Anki imports keep the media from the original deck where MemDecks can read it safely.
Where the extra lexical detail comes from
Part of speech, gender, distinct senses, example sentences and register-tagged synonyms are enrichment, produced by a language model in a pass separate from translation and reviewed against a structured contract before it is stored. It is optional by design: a card without enrichment is complete, it simply shows less. Enrichment never overrides the machine-translation baseline that made the card.