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.
In more detail
Laban and colleagues found that models answer worse when a task arrives over several turns than all at once, partly because early guesses stay in the history. Consolidate facts into one current statement rather than letting versions pile up.
Related terms
- 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.
- context compaction — Replacing older turns of a long conversation with a model-written summary so the history fits the context again.
Taught in
- Memory And Compaction — AI Engineering: LLM Apps, RAG & Agents
- When Context Fails — AI Engineering: LLM Apps, RAG & Agents
References
- Drew Breunig: How Long Contexts Fail (2025-06-22)
- Laban et al.: LLMs Get Lost In Multi-Turn Conversation (arXiv 2505.06120)