Reproduce “Agent Zero Memory: Provenance-Aware Long-Term Memory 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
Ming Wu, Pengyuan Zhu
The authors present an agent memory system that combines a timeline, a relationship graph, and source-backed durable facts. It routes queries to these stores and opens supporting evidence before answering. The system reports judged accuracy of 95.60% on LongMemEval and 93.60% on LoCoMo. Its retrieval ablation favors hybrid search, but cross-system comparisons are not controlled head-to-head trials.
Key insights
- The effectiveness of LLM agent memory is largely determined by its structured, provenance-aware architecture and retrieval discipline, rather than the specific LLM backbone's parametric knowledge.
- A hybrid retrieval approach, combining dense embeddings and lexical search, provides superior performance compared to either method in isolation.
- Maintaining provenance and enabling non-destructive updates for knowledge items are essential for temporal reasoning, auditing, and preventing fabrication.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.