context poisoning
A mistake or false claim entering the context, for example from a hallucination or a bad tool result, and being repeated as fact on later turns.
In more detail
Drew Breunig named it among four ways long contexts fail. Once a wrong fact sits in a goal, summary or memory, the model keeps building on it, so check facts before they enter durable state.
Related terms
- context distraction — A long context pulling a model towards repeating its history or attending to similar but irrelevant passages instead of the task.
- context confusion — Irrelevant material in the context, such as unneeded tool definitions, lowering the quality of a model's choices.
- context clash — Parts of the context contradicting each other, such as an early wrong attempt and a later correction, leaving the model unsure which to follow.
Taught in
- Memory And Compaction — AI Engineering: LLM Apps, RAG & Agents
- When Context Fails — AI Engineering: LLM Apps, RAG & Agents