---
title: "What is a knowledge update?"
description: "A knowledge update replaces a stored fact's value when the world changes, keeping predecessors distinguishable. How supersession works, with an example."
canonical: https://past.dev/glossary/knowledge-update
last-updated: 2026-09-02
---
# What is a knowledge update?

Source: https://past.dev/glossary/knowledge-update

A knowledge update is the replacement of a stored fact's value when the world changes, recorded so the current value and its predecessors stay distinguishable. When a budget moves from 40,000 to 55,000, an update stores the new value, closes the old one's validity, and links the two. The old value remains queryable as history instead of surfacing as a live answer.

## How a knowledge update is recorded

An update is a pair of operations on the same subject. The previous fact's [validity window](/glossary/validity-window) is closed at the change date. The new fact opens its own window and carries a [supersession](/glossary/fact-supersession) link back to what it replaced.

- **Delete is the wrong tool**: deleting the old value erases history and breaks answers about the past.
- **Overwrite is the wrong tool**: overwriting keeps one value with no record that it ever differed.
- **Supersession is the right tool**: both values persist, each with the dates it was true.

> **Distinction**
>
> An update records a change in the world. A correction records an error in the store. A revised budget really was the old value for a time; a mistyped budget never was.

## Why it matters for AI agents

Most facts an agent stores will change: owners, budgets, dates, preferences, org charts. A memory without an update mechanism accumulates every version as equally current, and recall becomes a guess between them.

With recorded updates, three queries stay answerable from one store: the current value, the value on a given date, and the change history. That is the model a [temporal knowledge graph](/glossary/temporal-knowledge-graph) implements. In past.dev, a changed fact closes the previous fact's validity window, and both values stay queryable with their dates.

## A knowledge update in practice

A project budget is revised twice during 2026. Each source message carries its own date, and each revision becomes an update.

| Date | Statement in source | Stored result |
| --- | --- | --- |
| January 9, 2026 | Kickoff email sets the Delta budget at 40,000 EUR | Fact opens: budget 40,000, valid from January 9 |
| March 12, 2026 | Steering meeting raises it to 55,000 EUR | Update: 40,000 closes at March 12; 55,000 opens, linked to it |
| June 2, 2026 | QBR cuts the budget to 48,000 EUR | Update: 55,000 closes; 48,000 opens as current |

In September 2026 the store answers all three shapes: current budget 48,000, February budget 40,000, and a three-step change history with dates and sources. [Facts that change over time](/guides/facts-that-change-over-time) treats the design in depth.

## Related concepts

- **[Fact supersession](/glossary/fact-supersession)**: the link from a replacing fact to the fact it replaced.
- **[Validity window](/glossary/validity-window)**: the dated span during which one value was true.
- **[Memory staleness](/glossary/memory-staleness)**: what accumulates when updates never happen.
- **[Point-in-time recall](/glossary/point-in-time-recall)**: querying the value as of a date.
- **[Temporal reasoning](/glossary/temporal-reasoning)**: answering current, historical, and ordering questions over updated facts.

## Frequently asked questions

### How do AI agents update their memory when a fact changes?

Temporal memory systems record the new value, close the previous value's validity period at the change date, and link the two. Both stay stored, so the agent answers with the current value and can still explain the history.

### Why keep old values instead of overwriting them?

Questions about the past, audits, and change tracking all need the old values. Overwriting also destroys the evidence needed to notice that a source was reporting an outdated figure.

### What is the difference between a knowledge update and a correction?

An update records that the world changed, so both values were true in their own periods. A correction records that the store was wrong, so the old value was never true. Systems that separate the two can tell a revised budget from a mistyped one.

## Related

- [Fact supersession](https://past.dev/glossary/fact-supersession)
- [Validity window](https://past.dev/glossary/validity-window)
- [Facts that change over time](https://past.dev/guides/facts-that-change-over-time)
- [Memory API quickstart](https://past.dev/docs/memory-api/quickstart)
- [Benchmark methodology](https://past.dev/benchmarks/methodology)