Reproduce “FadeMem: Biologically-Inspired Forgetting for Efficient 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
Lei Wei, Xiao Peng, Xu Dong, Niantao Xie, Bin Wang
FadeMem manages agent memories by assigning them different decay rates across long- and short-term layers, while using conflict resolution and memory fusion to update or consolidate information. In a simulated 30-day LTI-Bench evaluation, it retained 82.1% of critical facts using 55.0% of memory, versus Mem0’s 78.4% and 100%; storage is counted by memories, not tokens or physical size.
Key insights
- Active forgetting, as observed in human cognition, is an adaptive feature essential for preventing cognitive overload and maintaining information relevance in long-term AI agent memory.
- Dynamically assigning memory importance and applying adaptive, differential decay rates across memory layers improves efficiency over fixed-capacity or equal-importance retention strategies.
- Leveraging LLMs for conflict resolution and memory fusion provides robust management of new information and effective compression of redundant data, enhancing overall memory integrity.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.