People searching translation memory cleanup, remove duplicate TM entries, translation memory maintenance, clean stale translation memory, TM editor duplicates, or how to improve translation memory quality are dealing with a productivity problem that grows invisibly. A translation memory can accumulate years of good work and still become slower to use because old terminology, duplicate targets, aligned noise, outdated style and conflicting entries compete with current approved translations.
A fast translation workflow therefore includes periodic translation memory maintenance. Current localization guidance explicitly notes that TMs tend to collect entries that are no longer useful, including duplicate or inconsistent translations and terminology that changed on the customer side. CAT tools provide TM editors, duplicate filters, import controls, penalties, metadata and cleanup operations because translation memory is not a static archive. It is a working retrieval system.
This article owns one narrow reader job: remove or demote TM entries that create repeated review cost while preserving useful historical bilingual evidence. It is not a general explanation of translation memory, not a TM penalty article, and not a working-versus-master TM architecture page. Cleanup is the maintenance operation that asks: which stored translation units should still be allowed to influence current translators?
Quick answer
Clean a TM when it repeatedly produces:
- duplicate source/target pairs;
- competing targets for the same source/context;
- obsolete terminology;
- wrong-locale wording;
- bad alignment;
- unreviewed MT;
- malformed tags;
- stale project-specific text.
Use:
back up TM → define cleanup target → filter candidate entries → sample before deleting → merge/demote/correct → rebuild metadata if needed → test retrieval → monitor future matches
Do not delete history merely because it is old.
Remove or demote entries because they are misleading, redundant or no longer authoritative.
Why TM clutter costs translation time
A dirty TM creates micro-decisions.
Source segment appears.
Results pane shows:
- current approved translation;
- old translation;
- duplicate old translation;
- vendor variant;
- machine-generated variant.
Translator stops.
Which one?
They inspect dates.
They inspect metadata.
They ask the reviewer.
This can happen hundreds of times.
TM cleanup converts repeated human arbitration into one resource-maintenance decision.
Duplicate versus alternative translation
Not every repeated source with different target is a duplicate problem.
The same source can legitimately require multiple targets because of:
- context;
- client;
- product;
- locale;
- gender;
- formality.
A true duplicate is redundant under the TM’s design.
A competing target may be legitimate if metadata/context distinguishes scope.
Cleanup begins with classification.
Worked example 1: exact duplicate import
A TMX file is imported twice.
The TM now contains identical translation units with identical source and target.
No value added.
Possible costs:
- database size;
- noisy editor results;
- duplicate counts.
Safe cleanup:
- filter exact duplicates;
- retain one;
- preserve metadata if needed.
This is low-risk maintenance.
Worked example 2: same source, old terminology
Source:
service plan
Old target uses deprecated term A.
New target uses approved term B.
Both remain in TM.
Translator sees both.
If old terminology is truly obsolete:
- correct/delete old entry;
- or mark/demote legacy TM;
- update metadata.
Do not preserve obsolete competition merely because it is historical.
History can live in archive, not active production TM.
Worked example 3: legitimate contextual variation
Source:
Open
TM contains:
- target as verb “open file”;
- target as adjective “open status.”
Both legitimate.
Do not merge.
Instead ensure context/metadata can distinguish them.
Cleanup is not synonym extermination.
Worked example 4: misaligned corpus entry
Aligned legacy documents created TM entry:
Source:
Payment terms apply.
Target actually belongs to next sentence:
Contact your account manager.
This is dangerous.
Delete or repair.
A fuzzy match from this entry can look plausible enough to mislead.
Alignment-origin resources deserve sampling and penalties until validated.
TM age is a signal, not a verdict
Old entries can be excellent.
A 2018 legal clause may still be authoritative.
A 2026 MT import may be poor.
Use:
- approval;
- domain;
- terminology currency;
- project scope.
Do not clean by date alone.
Build a cleanup taxonomy
Classify candidate issues.
Duplicates
Identical or functionally redundant.
Conflicts
Same source/context, incompatible targets.
Stale terminology
Target no longer approved.
Stale style
Meaning correct, style outdated.
Wrong locale
Valid language, wrong regional variant.
Misalignment
Source and target do not correspond.
Structural damage
Tags/placeholders malformed.
Low-trust origin
Unreviewed MT or noisy import.
Different issue classes need different action.
Cleanup actions
Delete
Use when entry is wrong/redundant and has no useful historical role.
Correct
Use when source-target pair should remain but target/metadata is wrong.
Merge
Use when duplicates represent one concept/entry.
Demote
Move to lower-priority/reference TM or apply penalty.
Archive
Keep history outside active production.
Do not assume deletion is the only cleanup tool.
The backup rule
Before major TM cleanup:
- export TMX;
- snapshot resource;
- record date.
Cleanup is high leverage.
A mistaken bulk delete can remove years of knowledge.
Reversibility is cheap insurance.
Sample before you delete
Filter 10,000 duplicate candidates.
Do not delete immediately.
Sample:
- start;
- middle;
- end;
- different clients/domains.
Confirm pattern.
Then scale.
A filter can be broader than you expect.
Duplicate creation mechanisms
Duplicates often arise from:
- repeated TMX import;
- multiple-translations setting;
- different context handling;
- confirming non-text segments;
- merging TMs;
- vendor handoffs.
Understanding cause prevents recurrence.
Multiple translations setting
Some TMs intentionally allow multiple translations for same source.
This can be useful when context is weak or variants are legitimate.
It can also create noise.
Cleanup must respect TM design.
Do not force one target if project genuinely supports variants.
Context-aware TM design
Context can allow multiple source-identical entries without ambiguity.
If context is stable:
- keep variants;
- rely on context match.
If context irrelevant:
- duplicates may be removable.
Architecture determines cleanup.
Worked example 5: same source, two locales
Source:
Color
Targets:
- UK variant;
- US variant.
If TM mixes target locales, that is resource-scoping problem.
Better:
- separate locale TMs;
- or metadata/filters.
Cleanup may involve splitting resource, not deleting language variety.
Worked example 6: old brand name
Old target:
Control Center
New brand:
Operations Hub
If historical manuals still need old name, move legacy entries to archive/reference TM.
Current production TM should surface new brand first.
Historical truth and current authority can coexist in different resources.
Worked example 7: style migration
Old target is formal.
New style guide prefers plain language.
Old target is semantically correct.
Options:
- keep in legacy TM with penalty;
- bulk rewrite reviewed entries;
- create new master TM.
Delete only if old style has no value.
The right action depends on future reuse.
Worked example 8: MT contamination
A TM imported raw MT segments.
Reviewers later discover hallucinated details.
Do not leave them active because they are “translations.”
Segregate machine-origin entries.
Only reviewed content should become master reuse if that is project policy.
Metadata is cleanup infrastructure
Useful fields:
- client;
- project;
- domain;
- product;
- date;
- origin;
- reviewer;
- approval status.
Without metadata, good and bad history becomes hard to separate.
TM cleanup often reveals why metadata matters.
Clean metadata too
Entries may have:
- wrong client;
- missing domain;
- outdated product.
This can misroute priority.
Cleanup is not only source/target text.
It includes retrieval context.
Worked example 9: copied client field
TMX import assigns all entries to wrong client.
Matches are linguistically correct but appear in projects they should not.
Fix metadata or split resource.
Do not blame translators for irrelevant suggestions.
Empty/non-text entries
TMs can contain:
- numbers;
- dashes;
- tags;
- empty-like segments.
These can create useless matches.
Filter and remove if no value.
Be careful with structured content where such segments may matter.
Segment-length anomalies
Very long or very short entries can indicate alignment problems.
Sample:
- 200-word source vs 5-word target;
- 1-word source vs paragraph target.
Not proof, but useful anomaly filter.
Punctuation-only differences
Multiple entries differ only by punctuation.
Could be:
- legitimate style history;
- duplicate noise.
Normalize or merge if current policy makes distinction irrelevant.
Capitalization-only differences
Brand/style may make them meaningful.
Do not merge blindly.
Whitespace differences
Leading/trailing spaces may create duplicate-looking units.
Clean only if file format doesn’t require them.
TM text should reflect reusable linguistic unit.
Tags and TM cleanup
Stored tag patterns can differ across files.
Do not remove tagged variants simply because plain text matches.
A tag-rich source can legitimately require different target structure.
Use tag-aware comparison.
Source normalization
Some TMs normalize punctuation/case internally.
Understand editor’s duplicate filter.
Two visually different entries may already be equivalent under matching logic.
TM cleanup and penalties
If you cannot clean immediately:
- apply penalty;
- lower priority;
- mark reference-only.
This protects translators while maintenance continues.
Penalty is containment.
Cleanup is correction.
TM cleanup and thresholds
Lower-quality TM may still be useful below automation threshold.
Keep for manual research, exclude from pre-translation.
This is resource tiering.
TM cleanup and working/master architecture
Working TM can be noisy by design.
Master should be cleaner.
Periodic cleanup/promotion:
- review working entries;
- promote approved;
- archive/reject weak.
This prevents master degradation.
TM cleanup and concordance
Concordance searches all stored history.
Dirty TM makes concordance noisy.
Cleanup improves research quality, not only automatic matches.
TM cleanup and predictive typing
If TM feeds autocomplete fragments, stale phraseology can appear at cursor.
Cleaning improves suggestion quality.
TM cleanup and fragment assembly
Fragment systems can reuse subsegments from bad entries.
Misaligned TM can pollute phrase suggestions.
Another reason to remove structural noise.
TM cleanup and source updates
A clean TM makes fallback retranslation after source updates safer.
Dirty TM turns revision into ambiguity.
Maintenance improves future resilience.
Failure mode 1: bulk delete without backup
Catastrophic if filter wrong.
Backup first.
Failure mode 2: delete all conflicting translations
Some are contextually valid.
Investigate.
Failure mode 3: leave stale terminology because “reviewers can fix”
You are outsourcing known resource defect to every future reviewer.
Fix once.
Failure mode 4: clean only source/target, ignore metadata
Wrong-domain suggestions persist.
Resource still noisy.
Failure mode 5: merge locales
You create one target that is wrong for half the markets.
Keep locale boundaries.
Failure mode 6: use TM as archive and production simultaneously
Historical variants flood current suggestions.
Separate active from archive/reference.
Failure mode 7: delete old approved legal wording after stylistic rebrand
Future legacy contract may need it.
Archive rather than erase.
A cleanup workflow
- export backup;
- define objective;
- filter candidates;
- sample;
- classify issue;
- delete/correct/merge/demote;
- retest retrieval;
- monitor translators.
This is controlled maintenance.
Duplicate filter workflow
For exact duplicates:
- filter;
- compare metadata;
- retain richest/approved;
- remove redundant.
Do not accidentally discard approval notes.
Conflict filter workflow
Same source, different target:
- check context;
- check product/locale;
- check date;
- identify current authority.
Then:
- keep both with scope;
- or retire obsolete.
Stale-term workflow
Search deprecated target.
Review:
- current projects;
- master TM;
- legacy TM.
Correct active master.
Archive historical as appropriate.
Update termbase.
Misalignment workflow
Find suspicious length/context.
Compare original files if needed.
Delete incorrect pair.
Do not “fix” target without verifying source alignment.
A 100-entry audit
For a million-entry TM, full manual review impossible.
Sample 100:
- random;
- recent;
- high-use;
- duplicates;
- longest;
- shortest.
Estimate health.
Then target cleanup where needed.
Usage-driven cleanup
Which entries actually surface?
A bad entry never matched in five years may be low priority.
A bad top match repeated daily is urgent.
Focus maintenance where human cost occurs.
Reviewer feedback as cleanup signal
If reviewers repeatedly correct same old target:
- search TM;
- repair entry.
One reviewer correction should improve resource.
Translator skip-rate as signal
If translators always skip first match from TM X, resource is low-value.
Investigate.
Termbase/TM conflict
TM uses old term.
Termbase uses new.
Translator keeps editing TM matches.
Cleanup TM or migrate terminology.
Do not accept permanent friction.
TM migration after terminology change
For high-value stable phrases:
- search old term;
- update target entries;
- verify grammar.
Do not Replace All blindly across polysemous strings.
Controlled batch correction.
TM dedup after merge
When merging vendor TMs:
- normalize metadata;
- dedup;
- resolve conflicts.
Do not dump every resource into master without curation.
TM import policy
Before importing TMX:
- inspect origin;
- languages;
- locale;
- metadata;
- duplicate behavior;
- review status.
Prevention cheaper than cleanup.
TM export policy
When sharing cleaned master:
- document scope;
- date;
- locale;
- status.
Avoid reimporting outdated version later.
Naming TMs clearly
Use: Client_Product_Locale_Master_2026
not: TM1.
Resource names reduce accidental misuse.
Version TMs
For major cleanup:
- master_v1;
- master_v2;
- archive.
Avoid overwriting only copy with irreversible edits.
TM size and performance
Very large TMs can slow management and increase noise.
Size alone is not reason to delete good data.
Partition by:
- client;
- product;
- domain;
- era.
Retrieval quality improves.
When to create a new TM instead of clean
If resource has:
- pervasive misalignment;
- unknown provenance;
- mixed locales;
- unreviewed MT;
- inconsistent metadata;
it may be cheaper to build clean master from approved bilingual corpus.
Do not spend weeks polishing garbage.
When to keep a dirty TM as research corpus
Historical material may still be useful for concordance.
Mark:
- reference-only;
- penalized.
Keep out of pre-translation.
This preserves evidence without production risk.
TM cleanup and audit trail
For regulated work, record:
- what removed;
- why;
- who approved.
Do not destroy required records.
Governance comes first.
A quarterly cleanup routine
For active client:
- duplicate filter;
- deprecated term search;
- random sample;
- locale check;
- metadata spot check.
Small recurring maintenance prevents drift.
A project-end cleanup trigger
After major release:
- promote approved target;
- remove provisional;
- update terminology;
- dedup working TM.
This feeds clean knowledge into next release.
Measure cleanup value
Before:
- top-match acceptance;
- reviewer term corrections;
- duplicate count.
After:
- compare.
If translators use top match more often, cleanup worked.
The deeper principle: translation memory should reduce decisions, not create them
A TM is valuable because it remembers solved work.
When it stores conflicting, stale or wrong work without scope, it stops reducing decisions.
Every future translator must decide again.
Cleanup restores the core promise:
past approved translation should make present translation cheaper.
Advanced practice: clean the TM by retrieval consequence
Not every bad entry deserves equal urgency.
A useful cleanup order is:
Tier 1: entries that repeatedly surface as top matches
These directly waste translator attention.
Tier 2: entries that contain known wrong terminology
These can propagate systematic inconsistency.
Tier 3: entries with misalignment or factual corruption
These can produce misleading reuse.
Tier 4: low-use historical clutter
Useful to clean eventually, lower immediate cost.
This consequence-first approach prevents maintenance teams from spending days polishing entries nobody ever sees while known top-match problems remain active.
Build a “known bad entry” queue
Whenever a translator or reviewer finds a bad TM suggestion, do not only correct the current target.
Capture:
- source;
- bad target;
- TM/resource;
- reason;
- preferred correction.
Then schedule resource repair.
One live incident should become one maintenance task.
This is the resource equivalent of promoting a comment decision into durable project knowledge.
Worked example 10: top match is always old disclaimer
A frequently reused disclaimer changed last year.
Old TM target still surfaces first.
Every translator manually edits it.
The team already knows the problem.
The cheapest action is:
- correct or retire old entry;
- ensure current approved disclaimer outranks it.
Cleanup pays immediately.
Worked example 11: same source, three historical targets
Source:
Continue
TM shows three targets from:
- desktop app;
- mobile app;
- web portal.
They are all valid in context.
Do not delete two.
Instead:
- separate resources by product;
- improve context/metadata;
- use context IDs.
The problem is not bad translation.
It is insufficient scoping.
Worked example 12: vendor memory with mixed target locale
A vendor TM contains:
- Brazilian Portuguese;
- European Portuguese.
Project target is pt-BR.
Some exact source matches return pt-PT wording.
This is high-cost noise.
Options:
- split TM;
- filter metadata;
- rebuild locale-specific master.
Do not rely on translators to detect every regional mismatch manually.
Worked example 13: punctuation-only duplicates after style migration
Old target uses:
- straight quotes.
New target uses:
- curly quotes.
Do you need two entries?
If content otherwise same and new style is authoritative:
- normalize current master;
- archive old.
Do not let typography history create permanent competing exact matches.
Worked example 14: source typo preserved in TM
TM source:
sytem update
Current source:
system update
The old target is correct.
A fuzzy match may recover it.
Should you keep typo source entry?
Maybe not in master.
Correct the source side if resource editor allows and audit trace permits.
A clean TM should reflect reusable source reality.
Worked example 15: alignment created sentence-shift chain
An aligned document was off by one sentence for 50 entries.
This is not 50 isolated bad translations.
It is one alignment block problem.
Delete/re-align block from source files.
Systemic corruption needs systemic repair.
TM cleanup and alignment quality
When importing aligned bilingual corpora, record origin metadata.
Example:
Origin=Alignment_2023_manuals.
Later, if quality issue found, you can filter the entire imported cohort.
Without provenance, cleanup becomes forensic work.
TM cleanup and TMX imports
Before importing external TMX:
- create temporary staging TM;
- inspect;
- deduplicate;
- map metadata;
- then merge approved subset into production.
This is safer than importing directly into master.
Staging resources reduce contamination risk.
TM cleanup and reverse lookup
Some TMs support reverse lookup.
A source-target pair may be used in opposite direction.
Cleanup consequences therefore extend both ways.
A bad target can become bad source evidence later.
Know whether TM is reversible.
TM cleanup and multiple target languages
Some systems separate TMs by language pair.
Others support multilingual resources.
If multilingual:
- cleanup must preserve language field;
- do not merge identical source across target languages.
The unit of reuse includes language direction.
TM cleanup and context type
Context can be:
- previous/next segment;
- ID-based.
Two entries that look duplicate in source/target may have different context.
Do not delete until you understand context semantics.
A duplicate filter can overreport when context is intentionally distinct.
TM cleanup and IDs
Context IDs can change across source versions.
Old entries may become less useful without being wrong.
This is not necessarily deletion issue.
Their textual matches can still be valuable.
Demote by context confidence, not age.
TM cleanup and confirmation status
If TM metadata records:
- translator;
- reviewer;
- client approval;
use it.
Prefer entries with stronger approval.
When duplicate targets conflict, review status can break tie.
TM cleanup and timestamps
Recent is not automatically better.
But date can reveal:
- pre/post terminology change;
- pre/post rebrand.
Use date as one signal.
TM cleanup and user metadata
If one source of entries is systematically weak:
- training;
- review;
- resource segregation.
Do not silently clean forever without addressing origin.
Maintenance should reduce future defects.
TM cleanup and quality incidents
After delivered mistranslation traced to TM:
- remove/correct bad entry;
- search same target/source across TM;
- inspect current projects;
- update QA/terminology if systemic;
- document incident.
Contain propagation.
TM cleanup after client corrections
Client correction should trigger:
- current target fix;
- master TM fix;
- termbase/style update if relevant.
Otherwise same wrong match returns next project.
TM cleanup after legal change
Legal wording can be version-specific.
Do not overwrite historical clause if old contracts still require it.
Separate:
- current master;
- legal archive.
Scope is better than deletion.
TM cleanup after product rename
Similar:
- current product master uses new;
- legacy product archive keeps old.
Use metadata/product version.
TM cleanup after style change
If style change affects thousands of units, full TM rewrite may be expensive.
Options:
- new master seeded from final current project;
- penalty old TM;
- gradual migration.
Choose based on reuse volume.
TM cleanup and machine translation engines
Some workflows feed TM into adaptive MT or LLM retrieval.
Bad TM can contaminate generated output beyond direct matches.
Resource quality matters more when reused by AI.
Clean knowledge bases produce cleaner AI assistance.
TM cleanup and RAG-like retrieval
If translation systems retrieve similar bilingual examples for LLM prompts, duplicates and stale targets can overweight wrong wording.
Deduplication and currency become model-control steps.
This is a modern extension of classic TM hygiene.
TM cleanup and human trust
The results pane has a psychology.
If translators often see bad matches, they stop looking.
Then even good matches lose value.
Cleanup restores trust.
Trust is a productivity multiplier.
A “trusted top-10” audit
Pick ten common source sentences.
Inspect current top matches.
Ask:
- correct?
- current terminology?
- right locale?
- right style?
- right context?
If several fail, master TM needs attention.
This small audit is high signal.
A conflict-resolution matrix
For same source, different target:
Same context, one approved
Keep approved; retire other.
Different context
Keep both with context.
Different locale
Separate/scope.
Different era
Current vs archive.
Unknown provenance
Research or demote.
This prevents reflexive deletion.
A duplicate-retention rule
When exact duplicate entries differ only in metadata, decide which metadata you need.
Possible approach:
- merge metadata into one entry;
- retain strongest approval/date/source.
Do not discard provenance accidentally.
TM editor workflow
A productive editor session uses:
- filters;
- sorting;
- columns for metadata;
- batch actions.
Do not scroll million entries manually.
Maintenance is data work.
Filter ideas
- duplicate source/target;
- old term;
- specific creator;
- date range;
- origin;
- client;
- domain;
- target contains deprecated string;
- source/target length ratio.
Each filter should answer a cleanup question.
Sort ideas
Sort by:
- source;
- target;
- date;
- user;
- context.
Adjacent duplicates become visible.
Export candidate set before delete
For high-risk cleanup, export filtered candidates separately.
Now you have exact rollback for affected rows even if full TM backup large.
Cleanup in batches
Do not clean entire huge TM in one opaque operation.
Batch:
- duplicates;
- deprecated terms;
- misalignment;
- metadata.
Validate after each.
If issue appears, rollback smaller scope.
Post-cleanup retrieval test
Open representative project.
Search:
- known current term;
- known legacy term;
- context-sensitive string.
Confirm ranking improved.
Resource maintenance should be tested in actual translation workflow.
Post-cleanup pre-translation test
Run on sample file.
Check:
- fewer wrong exact matches;
- same good leverage;
- no lost critical reuse.
Cleanup should reduce noise without destroying coverage.
Monitor after release
Ask translators:
- better?
- missing useful old matches?
- still conflicting?
Human feedback validates maintenance.
Preventing duplicates at confirmation
Some CAT settings control multiple translations/context.
Review project settings.
If duplicates arise from confirmation behavior, fix configuration.
Cleanup without prevention is endless.
Preventing duplicate TMX import
Track imported file versions.
Do not import same TMX twice.
Use logs/naming.
Preventing raw MT contamination
Separate machine suggestions from human-approved TM.
Do not auto-confirm MT into master unless policy says.
Architecture prevents cleanup burden.
Preventing wrong-client contamination
Project templates should attach correct working TM.
A translator confirming Client A into Client B TM is preventable setup error.
Resource routing is quality control.
Preventing wrong-locale contamination
Name resources with locale and validate project target.
Automation can prevent accidental pt-PT into pt-BR.
Preventing stale terminology
When termbase changes:
- run TM search/migration on high-frequency old terms.
Proactive cleanup.
Maintenance cadence by project type
Continuous localization
Monthly/quarterly.
Annual report
After each release.
One-off project
Closeout cleanup enough.
High-regulatory
Change-controlled maintenance.
Do not impose one cadence universally.
Who should own TM cleanup?
Could be:
- lead translator;
- terminologist;
- localization engineer;
- PM.
Owner needs:
- language knowledge;
- resource authority;
- backup access.
Do not let nobody own it.
TM cleanup and confidentiality
TM may contain client content.
Do not send it to external cleanup tools without permission.
Backups also sensitive.
TM cleanup and retention
Privacy rules may require deleting old personal data from TM.
This becomes data-governance issue.
Follow policy.
TM cleanup and personal names
Named entities in TM can become stale or privacy-sensitive.
Do not treat TM only as linguistic asset.
It can store factual/client data.
TM cleanup and archive boundaries
Archive TMs can be read-only and excluded from automatic suggestions.
This is often best place for historical language.
Production master stays current.
The “one active truth” principle
For each current client/product/domain, there should be a clear active TM hierarchy.
Historical resources can exist.
But translators should know which resource is current truth.
Cleanup supports this clarity.
A 30-minute cleanup sprint
For a manageable TM:
- backup;
- remove exact duplicates;
- search top 10 deprecated terms;
- sample recent imports;
- test top matches.
Small recurring sprints can prevent large crises.
A large-scale cleanup project
For millions of entries:
- define data model;
- stage;
- automate safe duplicate detection;
- human-review ambiguous conflicts;
- rebuild master;
- validate.
Do not attempt manual row-by-row.
TM cleanup as infrastructure investment
Time spent cleaning is repaid across every future:
- translator;
- reviewer;
- release.
This is one of the rare translation tasks with compounding returns.
Final TM health checklist
- backup current;
- no known duplicate spam;
- current terminology;
- locale clean;
- metadata usable;
- aligned imports labeled;
- raw MT separated;
- top matches trusted.
If these are true, resource is doing its job.
Final cleanup stress test: challenge the resource from three directions
Before declaring maintenance complete, test the TM from three viewpoints.
Translator view
Open a representative current source sentence.
Does the top result help quickly?
Reviewer view
Search a known deprecated target.
Does it still surface as if current?
Project-manager view
Inspect metadata.
Can you tell:
- client;
- locale;
- product;
- approval state;
- origin?
A TM that passes all three is not merely “deduplicated.” It is operationally understandable.
Worked example 16: good match hidden below duplicates
Source:
Submit the request.
Results:
- duplicate old target A
- duplicate old target A
- current target B
Translator must scroll.
After cleanup:
- current target B
- legacy reference A lower-ranked elsewhere
Nothing about source text changed.
Human decision time changed.
This is the economic value of cleanup.
Worked example 17: wrong metadata creates wrong project match
Entry is current and linguistically excellent.
Metadata says:
Client B
but it belongs to Client A.
TM metadata prioritization pushes it into wrong project.
Cleanup fixes field.
Sometimes bad suggestion is not bad translation.
It is bad indexing.
Worked example 18: duplicate with richer context
Two identical source/target units.
One has:
- source file;
- context ID;
- reviewer approval.
One has none.
Keep richer record.
Do not randomly delete first/last duplicate.
Metadata quality should guide retention.
Worked example 19: corrupted tag pair in TM
Source tagged sentence was confirmed after manual tag damage.
Future fuzzy match brings broken tag structure.
Correct/remove entry.
Run tag QA in current projects.
TM can preserve structural defects just as efficiently as linguistic ones.
Worked example 20: target contains accidental source residue
Aligned TM target:
To restart, press reset botón.
Mixed source-language residue.
Search TM for source-language tokens can identify contamination.
This is especially useful after OCR/alignment/MT imports.
Search TM for QA patterns
You can run maintenance searches for:
- deprecated brand;
- forbidden term;
- double spaces;
- source-language residue;
- placeholder damage.
Resource QA mirrors project QA.
Why catch same error every release?
TM and spelling dictionaries
If reviewer corrects product spelling:
- TM entries may contain old misspelling;
- custom dictionary may still accept it.
Clean both.
Linguistic resources form an ecosystem.
TM and text expansion snippets
A snippet can insert current term while TM suggests old.
Translator sees conflict.
Resolve resource inconsistency.
Do not normalize human friction as normal.
TM and style cards
Style card changes:
use direct imperatives.
Old TM uses polite nominal instructions.
Decide:
- migrate high-use entries;
- penalize legacy.
TM should not fight current style on every segment.
TM and target-only review
Target-only pass may reveal patchwork caused by old TM.
If repeated stylistic seams trace to same resource, cleanup/demotion is more efficient than endless final polishing.
TM and source-version context
Old target can remain correct only for old source context.
If source keys reused improperly, TM can mislead.
Cleanup can flag suspicious context but source engineering may need fix.
TM health and exact-match rate
A high exact-match rate is not automatically success.
If translators edit most exact matches, the TM is low trust.
Measure unchanged acceptance, not only match percentage.
TM health and reviewer correction rate
If reviewers repeatedly alter exact matches:
- terminology stale;
- style stale;
- wrong scope.
This is powerful resource-quality evidence.
TM health and pre-translation debt
Pre-translation can fill 80% target.
If 40% requires heavy repair, coverage number is misleading.
Clean TM can reduce post-editing debt even if nominal coverage falls.
TM clean-room rebuild
For severely contaminated resource:
- choose approved final projects;
- align/export confirmed reviewer state;
- build new TM;
- test;
- keep old as archive.
Sometimes rebuilding is safest cleanup.
Preserve unique rare knowledge
Do not over-clean uncommon but correct technical wording simply because it appears once.
One future exact match may save major research.
Cleanup should remove low-quality noise, not rare expertise.
Preserve client-approved exceptions
A target may look inconsistent but be client-approved exception.
Metadata/note should explain.
Do not normalize away deliberate exception.
Document cleanup rules
For team:
- what counts duplicate;
- what archive;
- who approves delete.
This prevents each maintainer applying personal taste.
Final resource promotion rule
A master TM should contain content you are willing to show first to the next translator.
That is a strong practical standard.
If an entry would make you say:
“ignore that one,”
then active master is wrong place for it.
Final release checklist for cleaned TM
Before reattaching to live projects:
- backup retained;
- resource name/version updated;
- duplicate filter rerun;
- deprecated-term search rerun;
- sample pre-translation tested;
- top-match usefulness checked.
Then release.
Maintenance without validation is only editing data.
The final “would I trust this tomorrow?” test
A useful last question is brutally practical:
If this entry appeared as the top suggestion in a deadline-driven project tomorrow, would I want the translator to trust it?
If yes, it belongs in the active production TM.
If the answer is:
- “only if they know the old product,”
- “only if they ignore the old term,”
- “only if they realize this came from MT,”
- “only if they check the locale,”
then the resource needs better scoping, metadata, penalty, archive placement or deletion.
That test turns abstract database cleanliness into the human workflow the TM exists to support.
Maintenance should reduce future explanation
A clean TM requires fewer warnings from project managers such as:
“Ignore the first result.” “Use the second match, not the top one.” “That memory is old.” “Those exact matches are not really approved.”
Every such instruction is evidence that resource knowledge has not been encoded properly.
Cleanup is successful when those explanations disappear because the system itself reflects current authority.
One final release sample
Before putting the cleaned TM back into production, run a small sample project through it. Inspect at least one exact match, one fuzzy match, one context-sensitive short string and one known legacy phrase. The purpose is not to prove the entire database perfect. It is to make sure the cleanup changed retrieval in the intended direction without removing useful leverage.
If the sample is worse, restore or revise the cleanup. If the sample is better and the top suggestions are easier to trust, release the new TM version and retire the old production copy according to your governance policy.
Keep a quarantine TM for uncertain historical material
When you are unsure whether questionable entries should be deleted, move them into a clearly labeled quarantine or reference TM instead of leaving them in the active master.
A quarantine resource can preserve old client wording, uncertain alignment, pre-review MT, legacy terminology and disputed translations. Keep its priority low or exclude it from pre-translation. Translators can still search it deliberately when historical evidence is useful.
This is often safer than the false choice between “keep everything active” and “delete everything forever.” It separates retrieval authority from archival value while the organization decides what deserves long-term preservation.
Summary
Translation memory cleanup helps people translate quickly by removing or demoting entries that create repeated suggestion noise, stale terminology and conflicting reuse.
The reliable workflow is:
backup → classify duplicates/conflicts/stale/misaligned → sample → delete/correct/merge/demote → test retrieval → maintain regularly
Do not delete useful historical variation blindly.
Separate active authority from archive evidence.
A clean TM makes the first suggestion more trustworthy and every later project easier.
Frequently asked questions
Why do translation memories need maintenance?
Because duplicates, old terminology, wrong locale, misalignment and low-trust imports accumulate over time.
What is a duplicate TM entry?
A redundant translation unit under the TM’s context/multiple-translation design.
Should I delete every alternative target?
No. Different contexts or locales can justify multiple translations.
What should I do with stale terminology?
Correct active master entries or move legacy material to reference/archive resources.
Should I back up before cleanup?
Yes. Export or snapshot before high-leverage edits.
Can penalties replace cleanup?
They can contain low-trust resources temporarily, but known errors should be corrected or removed.
How do misaligned entries happen?
Alignment, bad imports or mismatched source/target files can pair unrelated segments.
Should raw MT be stored in master TM?
Only according to project policy; many workflows promote reviewed content only.
How often should I clean a TM?
Light recurring maintenance plus targeted cleanup after major imports, terminology changes or quality incidents.
How do I know cleanup worked?
Top matches become more useful, duplicate/conflict noise falls and reviewers make fewer repeated corrections.
Internal-link opportunities
- How People Translate Quickly | Working and Master Translation Memory
- How People Translate Quickly | TM Penalties and Match Thresholds
- How People Translate Quickly | TM Metadata Prioritization
- How People Translate Quickly | Translation Alignment
- How People Translate Quickly | Safe Find and Replace
- Master Art of Translation | The Translation Memory System