---
title: "What is data provenance?"
description: "Data provenance is the record of where information came from and how it changed: source, custody, transformations. Definition and why AI agents need it."
canonical: https://past.dev/glossary/data-provenance
last-updated: 2026-09-02
---
# What is data provenance?

Source: https://past.dev/glossary/data-provenance

Data provenance is the record of where a piece of information came from and how it reached its current form: its source, custody, and transformations. For an AI agent's memory, provenance means every stored fact can name the email, transcript, or ticket it was extracted from, with dates. Without it, an answer cannot be checked, a bad source cannot be traced, and a deletion request cannot be honored precisely.

## What a provenance record contains

- **Source**: the original record (message id, document, ticket) and who produced it.
- **Custody**: which systems held and passed the data, with timestamps at each step.
- **Transformations**: what was done to it: extraction, normalization, entity linking, summarization.

The W3C PROV recommendation standardizes this vocabulary as entities, activities, and agents; see the [PROV overview](https://www.w3.org/TR/prov-overview/).

Granularity matters. Provenance kept per document says which file a claim came from; provenance kept per fact says which sentence supports it. Verification and deletion both need the fact level.

## Why it matters for AI agents

A memory system transforms text aggressively: it extracts facts, merges duplicates, resolves entities. Each step is a place where an error can enter. Provenance is what makes the pipeline inspectable afterward.

- **Verification**: an answer that cites its source can be checked against it; an answer without provenance can only be trusted or ignored.
- **Poisoning response**: when a source turns out wrong or malicious, provenance identifies every fact it supported. See [memory poisoning](/glossary/memory-poisoning).
- **Deletion**: erasure requests remove a source and everything only it supported, which requires knowing what each source supports.

past.dev keeps attribution through this pipeline: recall returns evidence with its dates and sources, and DELETE /api/v1/ingest/{ingestionId} erases a source along with what only it supported.

## Data provenance in practice

An agent reports: "The Basel office closed in December 2024." A provenance-keeping memory can show the chain behind the claim.

1. Source: email from CFO Elena Marti, sent November 18, 2024, message id 8842.
2. Custody: ingested June 10, 2026 in the mailbox backfill, ingestion id b-1207.
3. Transformation: fact extracted June 10, 2026 with subject Basel office, value closing, event time November 18, 2024.
4. Entity link: "the Basel office" resolved to the office named in tickets and chat.

Each step is dated and attributable. The [data provenance guide](/guides/data-provenance) covers how to design and store this chain.

## Related concepts

- **[Audit trail](/glossary/audit-trail)**: the action-level record that per-fact provenance makes possible.
- **[Grounding](/glossary/grounding)**: tying answers to verifiable sources at response time.
- **[Ingestion time](/glossary/ingestion-time)**: when data entered the system, one field of custody.
- **[Fact extraction](/glossary/fact-extraction)**: the transformation step provenance must document.
- **[Memory poisoning](/glossary/memory-poisoning)**: the failure provenance helps trace and reverse.

## Frequently asked questions

### What is data provenance in simple terms?

It is the paper trail of a piece of information: where it came from, who handled it, and how it was changed along the way. With provenance you can trace any stored fact back to its original source.

### Why does AI agent memory need provenance?

Memory systems transform raw text into stored facts, and any step can introduce errors. Provenance lets an answer be verified, a bad source be traced and removed, and a deletion request be honored completely.

### What is the W3C PROV standard?

PROV is a W3C recommendation that models provenance as entities, activities, and agents, with defined relations between them. It gives systems a shared vocabulary for exchanging provenance records.

## Related

- [Grounding](https://past.dev/glossary/grounding)
- [Audit trail](https://past.dev/glossary/audit-trail)
- [Data provenance guide](https://past.dev/guides/data-provenance)
- [Memory API quickstart](https://past.dev/docs/memory-api/quickstart)
- [Benchmark methodology](https://past.dev/benchmarks/methodology)