---
title: "What is recall precision?"
description: "Recall precision is the share of returned memory that is relevant to the query. Why it matters for AI agents and what improves it."
canonical: https://past.dev/glossary/recall-precision
last-updated: 2026-09-02
---
# Recall precision

Source: https://past.dev/glossary/recall-precision

Recall precision is the share of returned memory that is actually relevant to the query; low precision buries the answer among near-matches. The complementary measure is coverage: whether the relevant records were returned at all. An agent's context window is small, so every irrelevant result returned displaces something useful. High recall precision means the agent reads a short list in which the needed records rank first.

## Precision and coverage

Two measures describe recall quality. Precision: of what was returned, how much is relevant. Coverage: of what is relevant, how much was returned. They trade off. Returning more candidates raises coverage and lowers precision; tight filters do the reverse. An agent needs both high at once: the relevant records present, and little else around them.

> **Terminology**
>
> Information retrieval calls this pair precision and recall. Because recall also names the memory operation on this site, these pages use coverage for the second measure.

## Why it matters for AI agents

An agent reads recall results inside a bounded context window, so every irrelevant record displaces a useful one and pushes the window toward [context rot](/glossary/context-rot). Agents also weight top-ranked results heavily. A near-match ranked first, an email about a similar customer or a superseded price, reads as the answer and produces a confidently wrong action. Low precision is therefore an action-quality problem before it is a search-quality problem.

## What improves recall precision

- **Exact identifier matching**: ids, names, and codes match literally alongside semantic search. See [hybrid retrieval](/glossary/hybrid-retrieval).
- **Entity resolution**: a query about Acme Corp returns records about Acme Corp under any spelling, and no records about similarly named companies. See [entity resolution](/glossary/entity-resolution).
- **Temporal filtering**: questions about October are answered from what was true in October. See [point-in-time recall](/glossary/point-in-time-recall).
- **Supersession awareness**: replaced facts rank behind their replacements. See [fact supersession](/glossary/fact-supersession).
- **Abstention**: when no record qualifies, the system says so rather than returning whatever ranks highest. See [abstention](/glossary/abstention).

Precision is measured against question sets with known relevant records. The [benchmark methodology](/benchmarks/methodology) describes the setup and scoring; testing guidance for one component is in [testing entity resolution](/guides/testing-entity-resolution).

## Related concepts

- **[Recall](/glossary/recall)**: the operation precision is measured on.
- **[Hybrid retrieval](/glossary/hybrid-retrieval)**: the main structural lever for precision.
- **[Context rot](/glossary/context-rot)**: the degradation low precision inflicts on the window.
- **[Evidence status](/glossary/evidence-status)**: how a system reports that nothing relevant was found.

## Frequently asked questions

### What is recall precision in AI memory?

It is the proportion of returned results that are actually relevant to the query. When precision is low the agent receives many near-matches and the correct record is easy to miss.

### Why does low precision hurt an AI agent?

The agent's context window is limited, so irrelevant results take the place of useful ones. Agents also tend to trust highly ranked results, so a wrong record ranked first can lead to a wrong action.

### How can recall precision be improved?

Combine semantic search with exact identifier matching, resolve entities so results belong to the right subject, filter by date, and rank superseded facts below current ones. Returning nothing is better than returning noise when no record qualifies.

## Related

- [Recall](https://past.dev/glossary/recall)
- [Hybrid retrieval](https://past.dev/glossary/hybrid-retrieval)
- [Abstention](https://past.dev/glossary/abstention)
- [Testing entity resolution](https://past.dev/guides/testing-entity-resolution)
- [Memory API quickstart](https://past.dev/docs/memory-api/quickstart)