Context compaction? I noticed llm seem to forget partially or completely the original task when context compaction happens. The problem is more serious with local llm with low context size.