context degradation

in English: context rot

advancedwidespread

Context degradation is the phenomenon where a language model works less accurately as more text accumulates in the conversation or task.

In plain terms

Imagine that papers are piled onto your desk all day and nothing is taken off. At 5 p.m., everything is still there, but the important note you received in the morning is barely findable, and you easily mix old, outdated notes with fresh ones. The same happens with a language model if too much text accumulates before it.

A model sees only a specific amount of text at once: this is the context window. In today's largest models, this can hold the equivalent of several books. Context degradation means performance does not drop suddenly upon reaching the limit but begins to weaken well before that, often unevenly and unpredictably.

Signs of degradation: the model forgets instructions given at the start of the conversation, mixes up details, repeats earlier mistakes, or references information that you have since overwritten. Therefore, a fresh, short conversation often yields better results than one dragging on for hours.

How it works

With every response, the model considers the entire preceding material: instructions, your questions, its own previous answers, and attached documents.
Simplified: the longer this material is, the more the model's attention is spread across text segments, so less is available for each important sentence.
Irrelevant, repetitive or contradictory parts act as noise: they distract the model, and similar but misplaced details are especially likely to mislead it.
Many models are worse at finding information in the middle of a text than at the beginning or the end.
Defense consists of maintaining context: starting a new conversation, summarizing previous ones concisely, and loading only the material truly necessary for the task.

An everyday example

You have been working with a chatbot for two hours on a grant application, and at the start you asked it to always use the formal 'you'. In the twentieth turn, it suddenly switches to the informal 'you' and brings back a paragraph you removed long ago. If you open a new conversation and paste the latest version with the instructions, it works precisely again.

Who is affected

employeeIf you work with long conversations or many attached documents, response quality may deteriorate unnoticed, and you must detect this. → Work & HR

developerA coding agent accumulates numerous files and logs over a long session, so maintaining context directly determines the reliability of its work. → Code

entrepreneur, executiveA customer service or internal assistant built on an AI agent can still make mistakes even if the model is good, when it receives too much disorganised material at once.

parent, student, teacherIf you upload an entire textbook for a single question, you may get a weaker answer than if you provide only the relevant chapter. → Education & Family

A common misconception

MythThe more background material I give the model, the better the answer I get, since everything fits into the large context window anyway.

In realityJust because something fits does not mean the model uses it well. Unnecessary text reduces accuracy, so well-selected, shorter material usually yields better results.

Limits and risks

Degradation is not necessarily uniform, and the model does not signal it: it responds with the same confidence even when it has lost the thread.
There is no general threshold where degradation begins; it varies by model and task, so during long work sessions, it is wise to check responses periodically.
Automatic summarization or compression helps, but important details may be lost in the process.
A larger context window alone does not solve the problem; it merely delays it.

Similar, but not the same

context window – The context window is the quantitative limit of text a model can view at once, while context degradation is a quality decline that occurs well before reaching that limit.
hallucination – A hallucination is a fabricated, false statement that can occur even in short conversations; context degradation is one possible trigger, but it can also cause other errors, such as forgetting instructions.
model drift – In model drift, the model's performance changes over months because the world or the data changes; in context degradation, this happens within a single conversation or task.

Related concepts

Broader

large language model

Narrower

lost-in-the-middle information

On the edge

context window, context engineering, context poisoning, token, hallucination, context compression, figyelmi mechanizmus

A brief history

The fact that language models use information in the middle of long texts less effectively was demonstrated by a 2023 study ('Lost in the Middle'). The term 'context rot' itself spread in 2025, mainly following the Chroma research team's report of the same name, which examined 18 models.

Where things stand

By the end of 2026, the term will be part of the everyday vocabulary of developers and teams building AI agents, and the phenomenon will be observable even in the latest models. Tools are increasingly handling this automatically, for example through compression, external memory, or distributing the task among multiple, clean-context agents.

Status: 2026 November

Also known as

context degradation · context degradation · context rot · context degradation · context degradation · context rot · context degradation

Sources

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.