RAG in a Trenchcoat (BRIDGE, 2611.22179, NeurIPS 2026)
TL;DR. Agent-memory research has stacked extraction, supersession, reflection and graph synthesis on top of plain retrieval. BRIDGE strips all of that out: chained observation notes, hybrid BM25 + cosine retrieval, chronological rendering, one LLM call per write or query, about 5-7K reader-prompt tokens per question. It ties the prior best within bootstrap confidence intervals on LongMemEval-S and BEAM and beats the best memory baseline on EverMemBench by 5.7-6.3 points at a matched reader. The authors' real point: benchmarks score what memory retrieves, not what it lets an agent do. On five adversarial probes where the agent must apply a remembered rule the question never mentions, no design wins everything. LangMem (atomic fact extraction) wins rule application, BRIDGE wins pattern induction, fact extraction loses the shape of a demonstrated pattern, and block summaries retrieve the rule but the agent often fails to apply it. Showing the rule more prominently closes most of that gap.
flowchart LR
O["Observations<br/><small>agent events</small>"] --> N["Chained notes<br/><small>one LLM call per write</small>"]
N --> R["Hybrid search<br/><small>BM25 + embeddings</small>"]
R --> C["Time-ordered context<br/><small>5-7K tokens</small>"]
C --> A["Agent<br/><small>answers or acts</small>"]
A -->|benchmark| P["Parity with SOTA<br/><small>LongMemEval, BEAM</small>"]
A -->|probes| F["Rule not applied<br/><small>retrieved but unused</small>"]
classDef input fill:#d0ebff,stroke:#1971c2,color:#1b1b1b,stroke-width:2px
classDef loop fill:#fff3bf,stroke:#f08c00,color:#1b1b1b,stroke-width:2px
classDef core fill:#e5dbff,stroke:#6741d9,color:#1b1b1b,stroke-width:2px
classDef exit fill:#d3f9d8,stroke:#2f9e44,color:#1b1b1b,stroke-width:2px
classDef err fill:#ffe3e3,stroke:#e03131,color:#1b1b1b,stroke-width:2px
class O input
class N,R loop
class C,A core
class P exit
class F err
How it relates to prior wiki pages
- Confirms the 11-17 memory update: Anthropic's study found a few strategy files beat 400+ archived notes, and MIRIX-2.0 kept complete records with read-time selection. BRIDGE is the cheapest point on that curve.
- Extends the 11-15/11-16 intervention-value thread (MemCredit, CoReMem: value a memory by whether removing it changes the next action): BRIDGE's probes are the benchmark version of that argument.
- Cost angle: one LLM call per write is the memory analogue of the 94.8% re-read finding: most agent memory spend is extraction and reflection that the benchmark does not reward.
Sources: arXiv · DAIR · @dair_ai (X Following feed, 2026-11-18)