A Practical Engineering Checklist for XTM-to-Phrase Translation Memory Migration
Plan an XTM-to-Phrase translation memory migration with immutable backups, pilot imports, leverage verification, and a rollback-ready cutover checklist.
- Cross-TMS failures concentrate in inline tags, segmentation, metadata, and encoding — verify each one, not just the import success message.
- Preserve an immutable source backup before any conversion, and audit export warnings without automatic tag stripping or reconstruction.
- Prove leverage in an isolated test TM with a representative pilot before touching production.
- Plan cutover and rollback in writing, and keep the source system readable until linguists verify leverage.
Moving translation memories from XTM Cloud to Phrase (Phrase TMS) is a controlled data transfer, not a file upload. A TMX export that completes without errors has still crossed four boundaries where systems legitimately disagree: how inline tags are represented, where sentences are split, which metadata survives, and how characters are encoded. Each disagreement is silent until leverage drops or a linguist reports mangled formatting. This guide is the engineering checklist for crossing those boundaries deliberately — with backups, pilots, and a rollback path — rather than discovering them in production.
Why cross-TMS TMX migrations fail
TMX promises interchange; practice adds implementation detail. Four causes recur.
- Inline tag representation. One system's paired begin/end tags become another's placeholder tokens. When paired-tag identifiers collide or a parser cannot map the structure, segments are rejected or tags are stripped — and stripped tags corrupt future matches on formatted content.
- Segmentation differences. Different segmentation rules split abbreviation lists, decimals, and headings differently, so a segment stored as one unit matches as a fraction in the new system. 100% context matches degrade to fuzzy without a single character changing.
- Metadata properties. Project codes, domains, client labels, and status live in custom properties in the export and in differently structured attributes in the destination. Without an explicit mapping, the provenance that drives leverage settings disappears.
- Character encoding. Export files that declare one encoding and contain another produce mojibake on accented and CJK content — often discoverable only by reading the records.
| Failure cause | Typical symptom | Verification check |
|---|---|---|
| Inline tags | Tag mismatch errors, or imports that quietly drop formatting | Count tags per segment in a source sample and in the imported sample; compare paired and placeholder representation |
| Segmentation | 100% matches become 70–80% fuzzies on identical text | Re-segment one known document in both systems and compare translation unit boundaries |
| Metadata | Missing client or domain context after import | Import a labeled sample and inspect how each custom property mapped |
| Encoding | Mojibake, rejected records, truncated fields | Open exported and imported files as text and verify accented and CJK samples survive exactly |
Preserving immutable source backups
Before any export or conversion, preserve an untouched copy of the source TM and its export files in an access-controlled location, and record the export date, source, language pairs, and settings selected. Everything else in this checklist assumes that copy exists and stays read-only: pilots, repairs, and re-imports all run against copies. Treat the backup as the definition of "the data we had" — if a dispute arises about whether the migration lost something, the answer is a diff against this copy, not a memory.
Verify the backup before you rely on it: open the exported files, confirm the expected language pairs are present, and spot-check record counts against what the source system reports. A backup that cannot be opened is a liability with a reassuring name, and discovering that during a dispute is the most expensive possible moment to find out.
Auditing export warnings without destructive fixes
Read the export log as an audit, not a formality. Warnings about malformed records, skipped units, or unexpected omissions describe exactly the segments that will misbehave later. Pause and investigate with the responsible platform support or a qualified migration specialist.
The line to hold: no automatic tag stripping, no tag reconstruction, no rewrite-in-place based on an assumed schema. Automated cleanup that removes or rebuilds inline markup does not repair data; it discards the formatting evidence future matches depend on, and it does so on the only copy that was still faithful. A text-only copy is fine as an additional reference — it is not a replacement for the structured original.
Warnings worth pausing on
Three categories justify stopping the line. Records rejected outright, because each one is a segment your linguists will silently lose. Format-related warnings on heavily formatted units, because those are the records where tag handling actually differs between systems. And counts that do not reconcile — fewer records exported than the TM reports, or more imported than exported — because arithmetic drift is how silent loss hides inside a successful-looking migration.
Running a pilot in an isolated test TM
Choose a representative sample: common and uncommon language pairs, short and long segments, heavily formatted content, placeholders and variables, and records that carry important metadata. Import it into an isolated test TM or another non-production destination approved for your account. Record which units were sampled and which import settings were used, so the pilot is repeatable and its results are arguable. A pilot that only tested your largest language pair tells you nothing about the CJK pair with heavy inline markup — sample for the failure modes in the table above, not for volume.
Recording the pilot
A pilot you cannot repeat is a pilot you cannot trust. Keep, with the test TM: the sample selection criteria, the export and import settings, the counts in and out, and the linguist verification notes. When someone later asks why the production import behaved differently, the pilot record is the control group that answers them — and if production matches the pilot, the record is the proof that process, not luck, carried the migration.
Verifying import logs and real leverage
Import summaries deserve skepticism. Read the destination's report, including rejected, skipped, and changed records, and reconcile counts against the source export before trusting the success line. Then verify what actually matters with the people who will use it: have linguists run real documents against the migrated TM and check that in-context and 100% matches appear where they did in the old system, especially on content with inline tags. Leverage checks belong to linguists because they can tell a true match from a plausible one — no import log can.
Planning cutover and rollback
Cutover is a scheduling exercise with an exit. Sequence production migration by language pair or business unit rather than all at once; agree a window when the source system is read-only or frozen for updates; and define, in writing, the rollback criteria — who decides, based on which evidence, that the migration reverts to the source TM. Keep the source system readable until verification completes; it is cheaper than any recovery plan. Communicate the freeze window and the decision point to linguists and project managers, because both sides touch the TMs during the window.
Where written engineering support fits
Migration problems concentrate in a handful of decisions — schema mapping, pilot design, rollback criteria — where a second pair of eyes with cross-TMS experience is worth more than another export attempt. If you want this checklist reviewed against your specific export, write to [email protected] with the subject "Translation memory migration inquiry"; support is written and asynchronous, so you can attach logs and samples directly. Prefer a web form instead of a desktop mail client? Email a migration inquiry. To see the surrounding toolchain, explore the BasetLoc tools for translation, preparation, and QC workflows. For the workflows that consume a healthy TM, read batch AI document translation with terminology workflows and preparing design and Office files for XLIFF localization.
Pre-cutover checklist
- Preserve an immutable, access-controlled backup of the source TM and exports, with date, source, and settings recorded.
- Audit the export warnings and resolve or document each one; no automatic tag stripping or reconstruction on the source.
- Map metadata properties explicitly, property by property, before the first import.
- Import a representative pilot sample into an isolated test TM and record the sample selection and settings.
- Reconcile import logs — rejected, skipped, and changed records — against source counts.
- Verify in-context and 100% leverage with linguists on real, formatted documents.
- Agree rollback criteria, decision owner, and freeze window in writing before production cutover.
- Keep the source system readable until verification closes; release it only after sign-off.
Technical FAQ
Can I clean the TMX with a script before importing?
Only as a copy, and only with full awareness that tag rewriting is data loss in disguise. Automatic tag stripping and reconstruction remove the formatting evidence that future matches depend on. Investigate warnings instead, on a copy, with the immutable backup untouched.
Why did matches drop even though the import reported success?
Import success means records landed, not that they match. Segmentation differences split stored units differently, and stripped or remapped tags prevent formatting-aware matches. Re-segment a known document in both systems and compare, then check formatted content with linguists.
Do we have to migrate every TM at once?
No, and you should not. Sequence by language pair or business unit, pilot each group, and keep the source readable until each group verifies. All-at-once cutover removes the rollback path exactly when you need it most.
How long should the old TMS stay available?
Until linguists have verified leverage on real documents in the new system and the rollback window has closed. The source TM is the reference copy that defines whether the migration lost anything; retire it by decision, not by deadline.
BasetLoc Engineering & Tools
Planning a localization workflow or TMS migration?
Explore our 13 commercial localization tools or send a written inquiry for custom migration and file-engineering requirements.