---
title: "What is a knowledge cutoff?"
description: "A knowledge cutoff is the date a model's training data ends. Why models miss newer facts and how retrieval and memory supply them, with dates and sources."
canonical: https://past.dev/glossary/knowledge-cutoff
last-updated: 2026-09-02
---
# What is a knowledge cutoff?

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

A knowledge cutoff is the date after which a model's training data stops; facts newer than the cutoff must come from retrieval or memory. A model trained on data through August 2025 can describe that world fluently and knows nothing reliable about March 2026. The gap is filled at inference time: retrieval, tools, and agent memory supply post-cutoff facts into the context window.

## Training knowledge and retrieved memory

| Property | Training knowledge | Retrieved memory |
| --- | --- | --- |
| Updated | Only by retraining | At every ingestion |
| Coverage | Public data before the cutoff date | Whatever the application ingested, from any date |
| Dating | Implicit; facts carry no dates | Explicit event time per fact |
| Attribution | None | Per-fact sources |
| Failure mode | Describes a past world with confidence | Gaps where nothing was ingested |

> **Boundary**
>
> Retraining moves the cutoff. It does not cover private data: a company's own email, tickets, and meetings never enter public training sets.

## Why it matters for AI agents

Most of what an agent is asked concerns post-cutoff reality: this quarter's budget, yesterday's email, the current owner of an account. Those answers cannot come from parameters at any cutoff, because they are private as well as recent.

The division of labor is stable across model upgrades. Parameters supply language, reasoning, and general world knowledge through the cutoff. Everything after it, and everything private, must arrive through the [context window](/glossary/context-window) from retrieval or [long-term memory](/glossary/long-term-memory). A memory API such as past.dev sits on that side of the division: it supplies post-cutoff facts at recall time with the dates and sources training data cannot carry.

## A knowledge cutoff in practice

A team's assistant uses a model with a mid-2025 cutoff. On March 9, 2026 the CFO renames the Delta project to Meridian in an email. Asked about Meridian in April, the model's parameters contain nothing: the name postdates the cutoff and never appeared in public data at all.

With the email ingested into memory, recall returns the rename as a dated fact (announced March 9, 2026, source: CFO email), the agent answers correctly, and the same store keeps answering after every model upgrade, whatever cutoff the new model has.

Two failure shapes deserve separation: a model answering about the post-cutoff world from stale parameters, and an agent whose memory is itself outdated. The second is [memory staleness](/glossary/memory-staleness); a fresher model does nothing to fix it.

## Related concepts

- **[Context window](/glossary/context-window)**: the channel post-cutoff facts arrive through.
- **[Long-term memory](/glossary/long-term-memory)**: the store that supplies them across sessions.
- **[Grounding](/glossary/grounding)**: attributing answers to sources instead of parameters.
- **[Memory staleness](/glossary/memory-staleness)**: outdated memory, the analogous failure on the retrieval side.
- **[Event time](/glossary/event-time)**: how supplied facts carry their dates.

## Frequently asked questions

### Why do AI models not know recent events?

Model weights are fixed when training ends, so anything that happened after the training cutoff is absent from them. Newer information must be provided at inference time through retrieval, tools, or memory.

### Can a knowledge cutoff be fixed by retraining?

Retraining moves the cutoff forward; it does not remove it, and it never covers private data such as a company's own email. Applications close the gap with retrieval and memory instead of waiting for new models.

### How does agent memory relate to the knowledge cutoff?

Memory holds facts the model was never trained on, both because they are newer than the cutoff and because they are private. At recall the facts arrive with dates and sources, which training knowledge lacks.

## Related

- [Long-term memory](https://past.dev/glossary/long-term-memory)
- [Cross-session memory](https://past.dev/glossary/cross-session-memory)
- [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)