Reproduce “Jev-Mem: System-One-Controlled Agentic Memory for Efficient AI 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
Dongming Jiang, Yi Li, Bingzhe Li
Jev-Mem uses a lightweight controller to handle structured memory construction and retrieval decisions, while a language model reasons over selected evidence and writes answers. On LoCoMo, its reported overall LLM-as-a-Judge score is 0.777 versus MAGMA’s 0.700. It reports 158-second memory building and 0.93-second query latency, but no ablation isolates the controller’s contribution.
Key insights
- Many memory operations require semantic judgments but not free-form language generation; replacing repeated LLM calls with bounded structured predictions can reduce overhead.
- Preserving all valid observations while applying selectivity to relation construction and retrieval avoids irreversible information loss during ingestion.
- Multi-relational graph traversal and adaptive stopping are designed to improve difficult multi-hop and adversarial queries while limiting irrelevant evidence and computation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.