Reproduce “Useful Memories Become Faulty When Continuously Updated by LLMs” 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
Dylan Zhang, Yanshan Lin, Zhengkun Wu, Yihang Sun, Bingxuan Li, Dianqi Li, Hao Peng
Repeated LLM-driven rewriting can make an agent’s long-term textual memory less reliable, even when every stored trajectory is successful or contains a ground-truth solution. Across interactive benchmarks and the ARC-AGI Stream, consolidation often improved performance initially but later caused sharp degradation; retaining episodic examples and limiting destructive updates proved more robust than continuously rewriting abstract lessons.
Key insights
- Memory utility is non-monotonic: consolidation may help early, but later updates can overwrite applicability conditions, merge incompatible task families, and reinforce incorrect procedures.
- The update schedule matters as much as the evidence itself. Streaming and heterogeneous batches perform worse than consolidating a well-grouped pool, with errors compounding when imperfect memories condition later rewrites.
- Concrete trajectories often preserve useful context that abstractions discard. Episodic-first or episodic-only storage is frequently competitive with, and sometimes better than, lesson-based memory.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.