Your agent's memory layer may be losing to plain retrieval

WHAT HAPPENED

A paper put naive RAG against complex agentic memory

The paper's baseline, Raw-RAG, uses dense retrieval and reranking over raw history. Across four benchmarks, it ranked first among the compared memory systems in six of seven benchmark-model settings.

THE NUMBER

6 of 7

settings where plain retrieval ranked first

Raw-RAG vs the compared memory systems, across four benchmarks.

THE REAL PROBLEM IN COMPANIES

What I see in companies: complexity gets added before it is tested

1. 01The team builds a memory layer first and never tests a plain baseline
2. 02Nobody measures whether the extra layer improves the answers
3. 03Summarizing or restructuring data can drop details the answer depends on

I'VE ALREADY SOLVED THIS

A building materials store in southern Brazil runs on simple retrieval

1. 0113,312 products in the ERP, 136 ways of writing the unit of measure
2. 02The AI never sees the catalog: a simple search hands 8 results to the model
3. 0365 out of 65 real customer questions found the right product

MY TAKE

Simple that works beats complex that impresses.

Before you add a complex layer, test a plain retrieval baseline and measure it. In production, the simplest retrieval that answers correctly is usually the one worth keeping. This paper points the same way.