---
title: 'Context rot'
description: "The degradation in an agent's ability to find, follow, or correctly weigh relevant information as its accumulated context becomes longer, noisier, stale, repetitive, or internally conflicting."
canonical_url: 'https://darkfactory.dev/glossary/context-rot'
markdown_url: 'https://darkfactory.dev/glossary/context-rot.md'
collection: glossary
date_published: '2026-08-05T00:00:00-04:00'
date_modified: '2026-08-26T00:00:00-04:00'
---

# Context rot


## Definition

Context rot is declining model or agent performance as usable signal is diluted by longer, noisier, stale, repetitive, or conflicting context. The input can still fit inside the advertised context window. The failure is that the system retrieves, weighs, or follows the important material less reliably.

The term has a narrow experimental use and a broader operational use. Chroma's July 2025 technical report varied input length across eighteen language models on controlled tasks and found nonuniform degradation as inputs grew. Position, semantic similarity, distractors, and document structure changed the results. Earlier work on information placed in the middle of long inputs had already measured a related positional failure.

In operating systems, context can also rot because its content ages. A steering file written for an older model may become redundant or harmful even if its token count stays fixed. Simon Boudrias reported that a Datadog team deleted stale AI steering material and improved an internal code-review evaluation. That is a self-reported case study without public data, but it supports periodically testing whether retained context still earns its place.

## Where the name came from

Chroma published "Context Rot" prominently in its July 2025 report and described the phrase as what the team was calling the observed degradation. Anthropic used the term in a September 2025 context-engineering article. These are clear records of current usage, but they do not prove that either organization was the first person or group to use the words.

## Distinguish it from nearby terms

- **Context-window overflow** means the requested input no longer fits. Context rot can occur well before that limit.
- **Context poisoning** introduces malicious or misleading material. Context rot can emerge from benign accumulation and staleness.
- **Compaction loss** occurs when a summary drops needed information. Compaction can also reduce rot by removing low-value history.
- A difficult later task is not evidence of context rot unless the comparison controls for task difficulty or otherwise isolates the context change.

## Operational significance

Evaluate context additions and deletions against representative tasks. Track length, source age, duplication, conflicting instructions, retrieval position, and task results. Prefer the smallest high-signal context that passes the evaluation. A larger window postpones truncation; it does not guarantee uniform use of every token.

## Check your understanding

An agent still receives a one-year-old setup rule after the product removed that requirement. The instruction fits and is retrieved correctly, yet following it makes the answer worse. That is operational context rot caused by staleness rather than overflow.

## Also called

context degradation

## Related terms

- [Context window](https://darkfactory.dev/glossary/context-window)
- [Context engineering](https://darkfactory.dev/glossary/context-engineering)
- [Prompt compression](https://darkfactory.dev/glossary/prompt-compression)
- [Working memory](https://darkfactory.dev/glossary/working-memory)
- [Context poisoning](https://darkfactory.dev/glossary/context-poisoning)

## Related factory areas

- [Context, memory, knowledge & skills](https://darkfactory.dev/factory/context-memory-skills)
- [Orchestration, state, concurrency & recovery](https://darkfactory.dev/factory/orchestration-state)

## Evidence and further reading

- [Context Rot: How Increasing Input Tokens Impacts LLM Performance](https://www.trychroma.com/research/context-rot)
- [Effective Context Engineering for AI Agents](https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents)
- [Lost in the Middle: How Language Models Use Long Contexts](https://arxiv.org/abs/2307.03172)
- [Datadog Deleted All Its AI Context. It Worked.](https://tessl.io/podcast/117/)
- [Long-Running Agents](https://addyosmani.com/blog/long-running-agents/)
- [Instruction Adherence in Coding Agent Configuration Files](https://arxiv.org/abs/2605.10039)