{
  "id": "context-rot",
  "code": "PTL-0169",
  "term": "Context Rot",
  "aliases": [
    "long-context degradation",
    "context degradation"
  ],
  "category": "failure-modes",
  "definition": "Context rot is the decline in a model's reliability as the amount of input in its context grows, even on simple tasks and well within its advertised context window.",
  "description": "A 2025 study of 18 models found performance degrading with input length in ways that depended on the task and on the content around the relevant information, such as similar-looking distractors. It is a practical reason to keep contexts lean, retrieve only what is needed, and compact long histories rather than relying on maximum context length.",
  "example": null,
  "broader": [],
  "narrower": [],
  "related": [
    "lost-in-the-middle",
    "needle-in-a-haystack",
    "context-window",
    "context-engineering",
    "context-compaction"
  ],
  "introduced": 2025,
  "added": "2026-10-10",
  "sources": [
    {
      "title": "Context Rot: How Increasing Input Tokens Impacts LLM Performance",
      "authors": "Hong et al.",
      "year": 2025,
      "url": "https://www.trychroma.com/research/context-rot"
    }
  ],
  "url": "https://protologue.com/t/context-rot/",
  "citation": "Protologue. (2026). Context Rot. In Protologue: A Taxonomy of Prompting and LLM Techniques (v1.4.0, PTL-0169). https://protologue.com/t/context-rot/"
}