A research agent offers to clean a Zotero library by merging records with similar titles. That sounds useful until the library contains a preprint, a later journal article and two copies of the same downloaded PDF.

Those are not necessarily the same cleanup problem. A repeated bibliographic record, a different version of a work and a repeated attachment require different decisions. Reducing the item count is not a sufficient definition of success.

For researchers delegating library maintenance, start with a proposed change list. The agent should explain what it believes is duplicated, what evidence supports that conclusion and what information must survive.

Methodology and intended audience

We reviewed Zotero's documentation on duplicate detection, related items and attachments on September 27, 2026. This is a proposed review workflow, not a runtime evaluation of a Zotero integration. No personal library, manuscript or attachment was opened or modified.

Our selection criterion was preserving scholarly context during cleanup. This is different from checking whether a DOI resolves or whether a paper supports a claim. The intended reader is a researcher or developer who wants automation without turning every similar record into the same source.

Separate three identities

Our recommended intake note distinguishes the bibliographic item, the version or manifestation of the work, and the attached file.

A bibliographic item carries citation metadata. A version distinction might reflect a conference presentation, an expanded article or another separately relevant form. A file is the actual artifact the researcher reads and annotates. Matching one identity does not automatically establish the others.

Consider a fictional pair of PDFs with the same filename. One could be an older download and the other a revised document. Conversely, two differently named files might contain the same bytes. Neither filename comparison settles whether two parent records should be merged.

Ask the agent to list these questions before proposing any deletion. If it only has metadata access, it must not claim to have checked the attachment contents.

Duplicate detection produces candidates

Zotero documents that duplicate detection uses fields including title, DOI and ISBN, with additional year and creator comparisons in relevant cases. Detection operates within a library rather than across separate group libraries. For genuine duplicates, Zotero recommends merging instead of deleting one record, preserving collections and tags and maintaining word-processor citation relationships. Zotero duplicate detection.

Treat the suggested group as a review candidate, not a command to merge everything selected. Our proposed plan records both item identities, conflicting metadata fields and the evidence for the intended common source.

For an approved merge, specify which field values should survive. Prefer the value supported by the actual source rather than assuming the most recently imported record is the most accurate. Preserve an unresolved field as a review question instead of inventing a correction.

Relate distinct versions instead of collapsing them

Zotero's Related Items feature can connect items within one library, and its documentation explicitly suggests connecting different versions of a work. Relations are reciprocal but not transitive: connecting A to B and B to C does not automatically connect A to C. Zotero related items.

Our recommendation is to retain separate records when the research task needs the distinction, and use a documented relationship rather than a merge to express their connection. For example, a review that compares an early claim with a later revision needs to know which version supported each observation.

A relationship is not a finding that the versions are substantively equivalent. Include a short note about why they are connected and which differences remain unexamined. Do not transfer an annotation's evidentiary meaning to another version without checking the corresponding passage.

Review attachments on their own terms

Zotero distinguishes standalone files from child attachments associated with bibliographic items. Standalone files lack metadata needed for features such as citation. Its documentation also distinguishes stored files managed by Zotero from linked files referenced at their existing location. Zotero attachment guidance.

Our proposed attachment inventory records the parent item, storage mode, file identity when available, accessibility and any annotation context that the workflow can actually inspect. Keep unavailable information marked unavailable.

Do not use library cleanup as implicit permission to delete original files from disk. Decide separately whether a repeated attachment should be retained, moved or removed, and identify what annotation or provenance information could be affected.

If a filename is the only matching evidence, defer the file decision. A tidier sidebar is not worth losing the artifact a researcher used to make a note.

Use a merge, relate or defer plan

Before applying changes, ask for a compact plan with one proposed action per candidate group:

Merge when the records describe the same intended bibliographic source and conflicting fields have an approved resolution.
Relate when separate records should remain but their connection is useful.
Defer when identity, version or attachment evidence is insufficient.

This is our decision framework, not a new Zotero confidence score. Add a reason and a reviewer to every proposed action. Review should focus on the most consequential ambiguity, not a cosmetic percentage assigned by a model.

Keep a pre-change inventory and follow the library's backup policy. Apply a small approved batch first, then inspect the resulting records and a representative document using their citations before continuing. These are suggested acceptance checks; no such batch was executed for this article.

A useful rehearsal and its limitations

Create a non-sensitive practice library with two imports of one source, two intentionally distinct versions and an unmatched PDF attachment. Write down the expected merge, relation and unresolved outcomes before running the agent.

The acceptance question is whether it preserves those distinctions and reports uncertainty, not whether it minimizes the final record count. Also ask it to explain which attachments it inspected and which it merely listed.

Metadata errors and incomplete access can still prevent a reliable decision. A clean bibliography does not establish the truth of the cited research. Our DOI and claim-support guide addresses that separate review.

Explore the skill directory for research workflows, then give the chosen agent a bounded cleanup plan. Library maintenance should make the evidence easier to trace, not silently simplify away the versions that explain it.