RAG in a Trenchcoat: Lightweight Agentic Memory (BRIDGE)
We show that a deliberately minimal RAG memory is enough on today’s agentic-memory benchmarks: BRIDGE, which combines chained observation notes, hybrid BM25/dense retrieval, and chronological context rendering, matches the leading systems on LongMemEval-S and BEAM and beats the best memory baseline on EverMemBench by 5.7–6.3 points. To measure what memory enables an agent to do rather than what it retrieves, we introduce five adversarial probes for implicit rule application, pattern induction, and multi-turn consistency. No architecture wins everywhere: atomic-fact extractors win rule application, while BRIDGE leads the strongest baseline by 33 points on procedural pattern induction.
RAG in a Trenchcoat: When Minimal Memory Is Enough for Agentic Systems, and When It Isn’t
Jingyu Liu, Zongze Li, Zach Xu, Zhanhui Zhou, Tahseen Rabbani, Dawn Song, Ce Zhang
Conference on Neural Information Processing Systems (NeurIPS 2026)
[Paper] [Code]
