---
title: "What is coreference resolution?"
description: "Coreference resolution defined: linking mentions inside one text to their referents, and why memory pipelines run it before entity resolution."
canonical: https://past.dev/glossary/coreference-resolution
last-updated: 2026-09-02
---
# What is coreference resolution?

Source: https://past.dev/glossary/coreference-resolution

Coreference resolution is determining which mentions inside a text refer to the same thing, such as linking a pronoun to the person it names. It operates within a single document or transcript, before any cross-record linking. When a meeting note says "Priya presented the roadmap and she owns the launch", coreference resolution attributes both statements to Priya, which gives later pipeline steps clean subjects to work with.

## How coreference resolution works

Coreference resolution runs on a single document, transcript, or message. It first detects mentions: names ("Priya Shah"), pronouns ("she"), and descriptions ("the account owner"). It then groups mentions that refer to the same thing into clusters, one cluster per referent.

The output is the same text with every mention assigned to a referent. Downstream steps no longer need to guess who "she" or "the client" is.

Transcripts are the hard case: speakers interrupt, names are misheard, and one meeting can use "she" across several participants. Precision matters more than coverage here, because a fact stored about the wrong person is worse than a mention left unresolved.

## Where it sits in a memory pipeline

Agent memory is built from text, and text refers to subjects indirectly. A sentence like "she approved the budget on June 9" is only useful if "she" is resolved to a named person before the fact is stored. [Fact extraction](/glossary/fact-extraction) therefore depends on coreference resolution: unresolved mentions produce facts about nobody, and wrongly resolved mentions produce facts about the wrong person.

Coreference errors are silent. The stored fact looks well formed; it is simply attributed to the wrong subject, and recall will repeat it with full confidence. The same failure surfaces in summaries and action items, where an unresolved "he" becomes an owner nobody can identify. The [benchmark methodology](/benchmarks/methodology) describes how memory systems are evaluated end to end, which is the level where attribution errors of this kind appear.

## Coreference vs entity and identity resolution

| Term | Scope | Links |
| --- | --- | --- |
| Coreference resolution | One text | Mentions to referents |
| [Entity resolution](/glossary/entity-resolution) | Records across systems | Records to real-world entities |
| [Identity resolution](/glossary/identity-resolution) | Channels and sessions | One person's identifiers to one profile |

The order matters. Coreference resolution runs first, inside each document. Entity resolution then links the cleaned documents' subjects across systems, and identity resolution maintains the person-level links. The [entity resolution guide](/guides/entity-resolution) covers the record-level step in depth.

## Related concepts

- **[Entity resolution](/glossary/entity-resolution)**: record-level linking across systems, the step after coreference.
- **[Identity resolution](/glossary/identity-resolution)**: person-level linking across channels.
- **[Fact extraction](/glossary/fact-extraction)**: the step that consumes resolved mentions and produces structured facts.
- **[Memory consolidation](/glossary/memory-consolidation)**: the background pass where these steps run after ingestion.

## Frequently asked questions

### What is coreference resolution in NLP?

It is the task of grouping mentions in a text that refer to the same thing, such as a name, a pronoun, and a description like the new hire. The output is a set of mention clusters, one per referent.

### What is the difference between coreference resolution and entity resolution?

Coreference resolution works inside one text and links mentions. Entity resolution works across records and systems and links whole records to a real-world entity. Pipelines usually run coreference first, then entity resolution.

### Why does coreference resolution matter for AI agents?

Agent memory stores facts extracted from text. If a pronoun is attributed to the wrong person, the memory records a false fact that recall will later repeat with confidence.

## Related

- [What is entity resolution?](https://past.dev/glossary/entity-resolution)
- [What is identity resolution?](https://past.dev/glossary/identity-resolution)
- [What is fact extraction?](https://past.dev/glossary/fact-extraction)
- [Entity resolution guide](https://past.dev/guides/entity-resolution)
- [Memory API quickstart](https://past.dev/docs/memory-api/quickstart)