Glossary and phrase memory

Glossa has two separate terminology tools. They solve different problems and work best together.

The distinction matters: the glossary is a constraint, while phrase memory is contextual support. The first tells the model what it must respect; the second shows approved examples that may help when the context is similar.

Glossary

The glossary is explicit and project-facing. You define source terms, required target terms, and optional notes, then Glossa injects those constraints into the run.

Use the glossary when:

• a term must always map to the same translation
• a project has house style or editorial vocabulary
• you want the judge to flag missing or incorrect terminology

Highlight legend

The Glossary tab in the Insight panel always shows the legend used in the document:

• blue underline: a glossary term found in the source text;
• green: the expected translation is present;
• pink: the expected translation is missing.

Search matches use yellow instead. You can change these colours under Settings → Highlights; hover a term to see its expected translation and notes.

Phrase memory

Phrase memory is retrieval-based. Glossa extracts reusable source-target pairs from approved work, stores them, and suggests them again for similar chunks later.

Use phrase memory when:

• recurring formulations appear across a corpus
• you want assistance beyond a small hand-maintained glossary
• the project evolves over time and should benefit from previously approved phrasing

How they differ

Tool	Best for	Input style
Glossary	Mandatory terminology	Manual entries
Phrase memory	Reusable local phrasing	Extracted approved pairs

Where each one acts in practice

• The glossary constrains the run up front.
• Phrase memory suggests reusable phrasing from prior approved work.
• The judge can still flag terminology or consistency failures after both.

Do not treat phrase memory as an absolute source of truth. A retrieved match can be perfect in one chapter and wrong in another: it should be selected, not merely accepted.

Recommended workflow

1. Start with a small glossary for names, terms, and non-negotiable translations.
2. Run a few chunks at a time: turn on the "Multiple chunks" toggle and use +/- to set a low number of segments to process.
3. Keep only the genuinely good outputs: lock those segments (the translation is final).
4. On a locked segment's Memory tab, extract phrases, review them (accept/discard/fix, or add manually), and confirm only when you're sure — nothing is ever saved automatically.
5. On later segments' References tab, check the suggested matches and only tick the relevant ones before running.

Good practice

• Keep the glossary short and strict; do not turn it into a full style guide.
• Use phrase memory for patterns, not for unquestioned automation.
• If a retrieved phrase is contextually wrong, reject it rather than weakening the threshold globally.
• Re-check terminology during audit, especially after refine or format stages.
• If a phrase is critical and always mandatory, promote it to the glossary instead of leaving it only in memory.