Reproduce “Remember the Decision, Not the Description: A Rate-Distortion Framework for Agent Memory” 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
Mingxi Zou, Zhihan Guo, Langzhang Liang, Zhuo Wang, Qifan Wang, Qingsong Wen, Irwin King, Lizhen Qu and 1 more
A decision-centric rate-distortion framework addresses the runtime memory bottleneck in LLM agents by quantifying memory quality based on its impact on downstream decision-making rather than descriptive fidelity. The proposed DeMem algorithm, which implements this framework, consistently improves agent performance across diverse long-horizon tasks, achieving higher overall scores on conversational benchmarks and more efficient use of memory resources.
Key insights
- A set of histories can safely share a single memory state if and only if there exists a common approximately optimal action for all histories within that state, establishing a decision-centric 'forgetting boundary'.
- Finding an optimal K-state partition for memory to minimize decision distortion is proven to be an NP-complete problem, necessitating efficient online learning approaches.
- The proposed decision covering and packing bounds formally relate the number of memory states to the achievable decision distortion, providing theoretical limits on memory efficiency.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.