agentsfactbearishSelf-summarization for LLM contexts comes at a significant performance cost, especially in human assistance domains where high quality data is scarceBAIR Blog29 Jul 2026http://bair.berkeley.edu/blog/2026/07/26/abbel/