Context Drift

AI & Agentic Systems • UX & Product Architecture

Definition

Context Drift is the gradual degradation of an AI agent’s reasoning coherence, instruction fidelity, or goal focus as its conversational context window fills with lengthy histories and tool responses.

Purpose

Pinpoints the root cause of late-session agent failures, guiding developers to implement periodic state summarization and context pruning mechanisms.

Key Concepts

Context saturationattention dilutioninstruction degradationstate decay.

How It Works

As token lengths expand, the model's self-attention attention weights disperse, reducing adherence to system prompts and causing task derailment.

Related Terms

Sources