Intermediate · Research

Context Rot

Visual diagram · (in preparation) · Math · (in preparation) · Worked example · 3 difficulty levels.

TL;DR. The measurable decay in a model's reasoning quality as its context window fills up — more input, worse answers.

Technical Definition

The measurable decay in a model's reasoning quality as its context window fills up — more input, worse answers.

How it works

Context rot names an empirical finding that reshaped how teams build agentic systems: model performance on a task degrades as the amount of irrelevant or merely longer context around it grows, even when the needed information stays present and the nominal context window has room to spare. Attention is a limited resource — with more tokens competing for it, recall gets mushier, instructions get followed less faithfully, and confabulation rates rise. The discovery pushed 2026 architecture toward context hygiene rather than ever-longer windows: aggressive retrieval curation, context compaction between agent steps, memory systems that summarize rather than accumulate, and eval suites that explicitly test long-context robustness. It is why 'maximum context' stopped being the headline spec and context management became an engineering discipline.

Related Concepts

• Agentic Context Management (ACM) — The discipline of deciding what an AI agent knows at each step — assembling, curating, compacting, and securing its working context.
• Agent Memory — Information an agent retains beyond the current context window — episodic history, learned facts, preferences, and task state.
• Chain-of-Thought (CoT) Prompting — Asking models to show step-by-step reasoning before giving a final answer, improving accuracy on complex tasks.
• Context Window — The maximum number of tokens a model can process in a single input, determining information capacity.