---
title: "What is recall in AI agents?"
description: "Recall is the operation that returns stored memory relevant to a query, ranked and dated. How AI agents recall facts with evidence and sources."
canonical: https://past.dev/glossary/recall
last-updated: 2026-10-17
---
# Recall

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

Recall is the operation that returns stored memory relevant to a query, ranked and, in evidence-based systems, dated and attributed to sources. It is the read half of an agent memory system: ingestion writes, recall reads. A recall request carries a natural language query; the response carries the records or facts that answer it, ordered by relevance, with enough context for the agent to use them and cite them.

## How recall works

Recall is the read operation of a memory system. Ingestion writes raw records in; recall brings relevant memory back out. A typical recall pass has four steps:

1. The agent sends a query in natural language describing what it needs.
2. The system gathers candidates by meaning and by exact identifiers. See [hybrid retrieval](/glossary/hybrid-retrieval).
3. Candidates are ranked by relevance to the query, with superseded facts ranked behind current ones.
4. The response contains ranked records with dates and sources. The consuming application can classify whether that evidence supports an answer; see [evidence status](/glossary/evidence-status).

## What a good recall response contains

A ranked list of text chunks is the minimum. An agent acting on the result needs more:

- **Ranking**: the most relevant records first, so a bounded context window gets the best few.
- **Dates**: the event time each record refers to, so the agent can tell current from historical. See [point-in-time recall](/glossary/point-in-time-recall).
- **Sources**: the record each answer came from, so claims can be verified. See [grounding](/glossary/grounding).
- **An empty result when nothing matches**: so the application can abstain instead of guessing. See [abstention](/glossary/abstention).

Recall quality is evaluated by asking questions with known answers against a stored history; the procedure is described in the [benchmark methodology](/benchmarks/methodology).

## Recall in practice

A support agent needs the history behind a ticket and asks what Acme Corp reported in the March 12 outage call. Recall returns the call record dated March 12 ranked first, the follow-up email dated March 14 second, and each carries the ticket and transcript it came from. The agent cites both dates in its reply. In [past.dev](/docs/memory-api/quickstart), recall is `POST /api/v1/recall` with a `query` field, and it returns ranked evidence with dates and sources.

## Related concepts

- **[Recall precision](/glossary/recall-precision)**: how much of what recall returns is actually relevant.
- **[Hybrid retrieval](/glossary/hybrid-retrieval)**: the candidate-gathering strategy behind good recall.
- **[Evidence status](/glossary/evidence-status)**: the label stating what the recalled evidence establishes.
- **[Point-in-time recall](/glossary/point-in-time-recall)**: recalling what was true as of a given date.
- **[Abstention](/glossary/abstention)**: recall declining to answer when evidence is absent.

## Frequently asked questions

### What does recall mean in AI agent memory?

Recall is the operation that fetches stored memory relevant to a query and returns it ranked, so the agent can use it in the current task. In evidence-based systems each result carries its date and source.

### How is recall different from search?

Search finds documents for a person to read. Recall serves an agent mid-task: it returns the specific facts or records the agent needs, ranked, dated, and attributed, so the agent can act on them directly.

### What should a recall response include?

Ranked results, the date each piece of evidence refers to, and its source. The consuming application assesses whether those results establish an answer. Missing dates or sources make that assessment harder.

## Related

- [Recall precision](https://past.dev/glossary/recall-precision)
- [Evidence status](https://past.dev/glossary/evidence-status)
- [Hybrid retrieval](https://past.dev/glossary/hybrid-retrieval)
- [Facts that change over time](https://past.dev/guides/facts-that-change-over-time)
- [Memory API quickstart](https://past.dev/docs/memory-api/quickstart)