Nightly Weight Consolidation Outperforms Cascading Compaction in LLM Memory Retention

A new study on arXiv (2605.24657) looks at two methods for keeping user-specific knowledge in large language models (LLMs) during inference-only use. The traditional method, called cascading compaction, only retains 36.8% of knowledge after three cycles. On the other hand, a new technique—nightly consolidation using reflection, synthesis, and Low-Rank Adaptation (LoRA) fine-tuning on one consumer GPU—boasts a much higher retention at 80.4%, which is a notable 43.6 percentage point gain. This research analyzed ten realistic software development conversations with 1,146 test questions across three memory types, finding the biggest improvements in procedural corrections, which saw a 36.3% rise. The results highlight the potential of weight-based updates for managing context-window limits and improving long-term user adaptation.