Minimal RAG memory matches complex agentic memory on benchmarks; no design wins everywhere
_reachsumit · x · 2026-11-18
Jingyu Liu, Dawn Song, Ce Zhang et al. publish "RAG in a Trenchcoat" on arXiv, a systematic look at agentic memory.
- Almost none of the existing memory primitives (extraction, supersession, reflection, graph synthesis) earn their cost on tested benchmarks. BRIDGE — a deliberately minimal RAG memory with chained observation notes, hybrid BM25+cosine retrieval, and chronological rendering — reaches parity with prior SOTA on LongMemEval-S and BEAM and beats the best baseline by 5.7-6.3 points on EverMemBench at matched reader, using 5-7K reader tokens per query and a single LLM call.
- The authors argue the parity itself is the finding: standard benchmarks score what memory retrieves, not what it enables when the agent acts. They introduce five adversarial probes judged at generation time via a cross-vendor verification chain.
- Across 5 probes x 20 scenarios, no architecture wins universally: atomic fact extractors (LangMem) win rule application, while BRIDGE's block-summary memory wins pattern induction.
More from coding & agent
- Microsoft's Autopilot Agent Is Built on OpenClaw, Now in Preview for Frontier Customers — pswider · 2026-11-18
- SimplePost Gives ChatGPT and Claude a Publishing Button for 10 Social Networks — haltakov · 2026-11-18
- One set of Obsidian notes feeds Claude Code, Codex and Gemini alike — dSebastien · 2026-11-18
- Obsidian Starter Kit gives Claude Code, Codex and Gemini the same persistent context — dSebastien · 2026-11-18
- AI tools start from zero at every prompt — the missing piece is persistent context — dSebastien · 2026-11-18
- 27 vibe-coded Finder clones appeared in a month, most already dead — KaKi_87 · 2026-11-18