A translation memory (TM) is a database of source sentences paired with their translations, which a CAT tool searches to suggest earlier translations when the same or similar text appears again.
Each entry is a translation unit: one source segment, its translation, the two language codes and usually some metadata, such as who created it and when. Every time the translator confirms a segment, the CAT tool saves it to the TM. On the next job, each new segment is compared with the stored ones:
A TM stores sentences, not terms. Approved terminology lives in a separate termbase.
Because reused sentences take less work, translators usually charge a share of the full rate for matched words. A CAT analysis counts how many words fall in each band, which is how a TM lowers the cost of updates and repetitive texts. A TM also keeps wording consistent across documents, translators and years.
Every CAT tool has its own TM format, such as Trados's .sdltm. To move a memory between tools, you export it as TMX, the standard exchange format. Our post on XLIFF, TMX and TBX explains how the formats fit together.
Last year you translated "Press the red button to stop the machine." and saved it to your TM. This year's manual says "Press the green button to stop the machine." Your CAT tool shows last year's translation as a fuzzy match of about 93%, highlights the changed word, and you only replace the colour.
A TM is only as good as what goes into it. Duplicates, empty units and conflicting translations produce bad suggestions; see how to clean a TMX file. TraduQ's free TM cleaner, TMX editor and TMX merge work in your browser, and your memory is never uploaded.