← News

Agents & MCPSources: arXiv · 18 Nov 2026

Minimal RAG memory matches complex agent memory systems in paper's tests

A preprint introduces BRIDGE, a deliberately simple retrieval memory, and reports parity with prior state-of-the-art on standard benchmarks while arguing that current tests miss what memory enables agents to do.

In plain words

Many AI agents use elaborate memory systems to remember past interactions. The paper's authors built BRIDGE, a very simple memory that just stores notes and retrieves them with standard search methods. In their tests it performed as well as far more complex systems, leading them to argue that benchmarks measure what memory retrieves rather than what an agent can actually do with it.

What was announced

The authors, writing in an arXiv preprint, describe BRIDGE as a minimal retrieval-augmented generation memory combining chained observation notes, hybrid BM25 and cosine retrieval, and chronological rendering. They report it reaches parity within bootstrap confidence intervals with the prior state of the art on LongMemEval-S and BEAM, and beats the best memory baseline on EverMemBench by 5.7-6.3 points at a matched reader, while running at roughly 5-7K reader-prompt tokens per query with a single LLM call per ingestion or query. They argue that standard benchmarks score what memory retrieves, not what memory enables when the agent acts.

To probe that gap, the paper introduces five adversarial tasks, judged at generation time by a cross-vendor verification chain, where a query can only be solved by applying a remembered rule or pattern that the query never states. Across 100 scenario runs, no architecture won universally: atomic-fact extractors such as LangMem did better on rule application, while BRIDGE's block-summary memory did better on pattern induction. The authors attribute this to atomic-fact extractors discarding the form a demonstrated pattern takes, while block-summary memory retrieves the rule but the reader does not act on it, a gap they say presenting the rule more prominently largely closes. The finding held across four domains, three graders, and three reader families.

What it does not show

  • The evidence is the paper's abstract only; the full methodology, baselines, and results have not been independently verified.

Sources

Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 19 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.