Reproduce “What Deserves Memory: Adaptive Memory Distillation for LLM Agents” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Wenquan Ma, Jiayan Nan, Wenlong Wu, Yize Chen
Researchers from Fudan University, Shanda Group, and Beihang University developed NEMORI, an adaptive, training-free framework for LLM agent memory distillation that leverages prediction error to identify and retain useful information. The system achieved up to 19.1% higher performance than baselines while reducing LLM calls by 59.5% and token consumption by 88% during response generation, also demonstrating enhanced scalability for long contexts.
Key insights
- Memory distillation can be effectively guided by a data-driven assessment of an experience's future utility, specifically by identifying information that the agent fails to predict based on its current knowledge.
- Structuring raw interactions into coherent 'episodes' and transforming them into narrative representations improves the efficiency and accuracy of memory formation and subsequent retrieval.
- A management-agnostic distillation layer can provide significant storage reduction (45-64%) and maintain or improve performance when integrated with existing third-party memory systems.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.