Minimal RAG memory matches SOTA agent memory systems, NeurIPS 2026 paper finds

dair_ai · x · 2026-11-19

A NeurIPS 2026 paper ("RAG in a Trenchcoat") shows that the complex machinery in agent memory systems—fact extraction, reflection, graph synthesis—rarely earns its cost on standard benchmarks.

The authors present BRIDGE: it stores chained observation notes, retrieves them with hybrid BM25 + embeddings in chronological order, and uses one LLM call per write/query (5-7K tokens per question). It reaches parity with prior SOTA on LongMemEval-S and BEAM, and beats the best memory baseline on EverMemBench by 5.7-6.3 points.

Crucially, the authors add five adversarial probes testing what memory lets an agent do—applying remembered rules or patterns the query never mentions. No single design wins everywhere: atomic-fact extraction (LangMem) wins rule application, while BRIDGE's block summaries win pattern induction. The takeaway: benchmarks score what memory retrieves, not what it enables agents to do—and the latter is what agentic memory should be evaluated on.

Related event: Minimalist RAG Memory Matches Complex Agent Memory Systems, Paper Finds(2 posts)→

Original post →

More from coding & agent

coding & agent channel →