Translation memory (TM)
A searchable bank of previously approved source and translated sentences, used to reuse exact wording and suggest close matches.
What it is
A translation memory is a database of source segments paired with their approved translations. When a new or changed string arrives, it offers exact matches and fuzzy matches from earlier work, so translators can reuse consistent wording instead of starting over.
Reach for a TM when a product has recurring UI, documentation, support copy, or releases across many locales. Trados built the commercial translation industry around this idea in the early 1990s, and the leverage report it produces still drives how vendors quote a job. Translation management systems usually update the memory as work is approved, and TMX is the common exchange format for moving one between tools, which matters on the day you switch vendors.
Gotcha: a matching sentence can mean something different on another screen. Keep locale, product, domain, and context metadata, and require human review for fuzzy matches. A TM remembers whole translated segments; a terminology database governs preferred terms inside them.
Ask AI for it
Build a translation-memory import and match service around TMX 1.4b. Store source locale, target locale, source segment, target segment, context, product, approval status, and timestamps. Return exact matches first, then rank same-locale fuzzy candidates with normalized Levenshtein distance. Protect ICU MessageFormat placeholders and inline tags during comparison, never auto-apply a fuzzy match, and write only human-approved translations back into the memory.