Translation memory
A database of previously translated segment pairs (source text and its approved human translation) used to suggest matches when the same or similar content recurs.
The honest version
Translation memory is the oldest AI in localization, and it still works.
The core logic is simple: if you translated “Click Save to confirm your changes” last year and a human reviewer approved that translation, you should not have to translate it again this year. TM stores that pair and surfaces it the next time the source appears. 100% matches require no translation. Fuzzy matches (segments that are similar but not identical) require review and light editing. The leverage compounds over time: the longer you maintain a TM, the more institutional knowledge it contains.
Fuzzy matching is scored by edit distance, typically expressed as a percentage. An 85% match means the strings share 85% of their characters in a defined comparison. Most organizations have rate schedules tied to match bands: 100% (no charge or minimal review fee), 95–99%, 85–94%, below 85% (full translation rate). These rates are where the economics of localization live.
Why it matters for translation
TM leverage is still the primary cost reduction mechanism in professional localization. For software with repetitive UI strings, changelogs, and documentation, TM hit rates of 40–70% are common. That percentage of content does not require translation, only review at reduced rates. In a well-maintained, domain-specific TM, every new release pays back part of the investment made in every previous release.
TM also enforces consistency automatically. If “Submit order” was approved as “Bestellung aufgeben” in German, that mapping recurs without human judgment every time it appears. No one has to remember. No one has to look it up. The TM is the institutional memory.
For LLM-based pipelines, TM is a primary retrieval source. A RAG system that retrieves approved segments from TM and injects them into the prompt gives the model access to your organization’s specific translation decisions, not just general language knowledge.
Where it fails
TM becomes a liability when it is large and inconsistent.
Multiple translators working over multiple years, with multiple style guides and varying quality standards, produce a TM where a given source segment might have five different approved translations across different projects. When the TM pre-translates a new file, it surfaces one of them, but which one, and is it still current? The translator or post-editor must evaluate not just whether the match is relevant, but whether the translation it proposes is still correct by current standards.
A polluted TM is worse than no TM. It introduces errors with the appearance of approved authority. Post-editors trust high-match-rate segments more than they should. Quality problems propagate.
TM is also segment-level by design. It has no understanding of context that spans multiple segments. A sentence whose translation depends on what came before it (pronoun agreement, topic continuity, register set by an earlier exchange) is invisible to TM matching. The match score tells you how similar the source string is; it says nothing about whether the proposed translation is contextually appropriate.
Finally: TM is a snapshot. It does not automatically update when the source changes. An approved segment from 2018 that reflects a deprecated product name will score 100% and be wrong.