Context Decay
Context decay is the failure mode where long conversations or long-running AI workflows lose, blur, or misapply earlier information. In Episode 17: 向量模型工程师：AI 的隐藏瓶颈与新时代的信息迷宫, 汉洋 / Han Yang and N 同学 / N Student discuss this as “context rot”: as multi-turn interaction grows, early state can become vague even when it technically remains in the window.
The concept qualifies simple long-context optimism. Longer windows help with some single-document tasks, but they do not automatically solve Retrieval-Augmented Generation, Deep Research, or agent work. Useful systems still need Context Engineering, retrieval, summaries, tests, source links, and Human Judgment Under AI to decide which information remains active.
Key Claims
- More context is not the same as better use of context.
- Long conversations can preserve tokens while losing operational clarity.
- Context decay can make an agent repeat work, forget constraints, undo earlier fixes, or answer from stale assumptions.
- Long-Horizon AI requires active state management: retrieve, summarize, forget, and re-check.
- RAG and knowledge graphs can be viewed as structured responses to context decay, but they need evaluation and upkeep.
Connections
- Context Engineering - practices for preserving useful context.
- Long-Horizon AI - model and harness direction that tries to handle extended tasks.
- Retrieval-Augmented Generation and Document Chunking - retrieval-based alternative to dumping everything into one prompt.
- Agent Harness and Persistent Agent Memory - system layers that decide what state survives.
- AI Coding Verification - catches regressions when coding agents lose prior context.