---
title: "What is grounding in AI agents?"
description: "Grounding ties an AI agent's answers to verifiable sources, so every claim traces to dated evidence. How grounded recall works in practice."
canonical: https://past.dev/glossary/grounding
last-updated: 2026-09-02
---
# Grounding

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

Grounding is tying a model's answer to verifiable sources, so each claim can be traced to the evidence that supports it. A grounded answer cites records that exist: a dated email, a transcript line, a stored fact with provenance. An ungrounded answer may still be correct, but no one can check it. For AI agents grounding is what separates memory from plausible text generation.

## What a grounded answer looks like

Grounding is a property of the whole path from storage to answer. The stored record keeps its origin. Retrieval returns the record with that origin attached. The answer cites the records it used. Concretely, a grounded answer carries:

- **A claim-to-evidence mapping**: each claim in the answer points at one or more records that support it.
- **Dates**: the event time of each piece of evidence, so the claim is anchored in time.
- **Sources**: the origin of each record and how it reached its current form. See [data provenance](/glossary/data-provenance).
- **Currency**: whether the evidence is current or superseded, so an old fact is never cited as the present.

## Why it matters for AI agents

A language model produces fluent text with or without support; fluency signals nothing about truth. Grounding constrains the agent's claims to what its evidence establishes and makes every claim checkable afterward. It also defines correct behavior when evidence is missing: the grounded response is [abstention](/glossary/abstention), stated through an explicit [evidence status](/glossary/evidence-status) rather than a fabricated answer.

Grounded answering is evaluated by checking claims against the returned evidence; the [benchmark methodology](/benchmarks/methodology) describes the checks. [past.dev](/docs/memory-api/quickstart) returns recall results as dated evidence attributed to sources, so an agent's answer can cite the records behind it.

## Grounding in practice

A support agent tells a colleague that Acme Corp is on the Enterprise plan. Grounded, that claim traces to a stored fact, the fact to its evidence, and the evidence to a source: the order confirmation email of May 3. When Acme disputes the plan in November, the agent produces the May 3 email and the dispute ends at the source.

Ungrounded, the same sentence is an assertion with nothing behind it. Checking it means redoing the research, and if the underlying data was wrong there is no chain to inspect. Building and keeping that chain is covered in the [data provenance guide](/guides/data-provenance).

## Related concepts

- **[Data provenance](/glossary/data-provenance)**: the source-and-custody record grounding depends on.
- **[Evidence status](/glossary/evidence-status)**: the machine-readable form of a grounding claim.
- **[Abstention](/glossary/abstention)**: the grounded behavior when evidence is absent.
- **[Audit trail](/glossary/audit-trail)**: grounded answers accumulated into an explainable history.

## Frequently asked questions

### What does grounding mean in AI?

Grounding means an answer is tied to verifiable sources, so each claim can be traced to a record that supports it. A grounded system can show its evidence; an ungrounded one can only repeat its claim.

### How does grounding reduce hallucinations?

Hallucination is fluent text without support. When answers must cite stored evidence and the system can decline when evidence is absent, unsupported claims are caught before they reach the user.

### What is the difference between grounding and provenance?

Provenance is the record of where a piece of information came from and how it was transformed. Grounding uses that record: it ties each claim in an answer back to sourced evidence.

## Related

- [Data provenance](https://past.dev/glossary/data-provenance)
- [Evidence status](https://past.dev/glossary/evidence-status)
- [Abstention](https://past.dev/glossary/abstention)
- [Data provenance guide](https://past.dev/guides/data-provenance)
- [Memory API quickstart](https://past.dev/docs/memory-api/quickstart)