Verbatim Chunks Beat Extracted Artifacts: A Controlled Ablation of Memory Representations for Long LLM Conversations

The study presents a controlled ablation comparing the effectiveness of verbatim conversation chunks versus LLM-extracted structured artifacts in conversational-memory systems. Results show that verbatim chunks significantly outperform extracted artifacts, achieving 43.9% vs. 28.0% on LoCoMo and 67.4% vs. 45.4% on LongMemEval-S. This indicates that retaining verbatim details is crucial for retrieval accuracy, suggesting that structured memory should complement rather than replace raw text in LLM applications.