Recent work shows that simple keyword matching can outpace semantic search in large‑language‑model agents. Researchers compared grep and vector retrieval on a 116‑question set from LongMemEval, using the Chronos harness and native CLI tools from Claude Code, Codex, and Gemini.

Across all harnesses, grep delivered higher accuracy, yet overall performance still hinged on harness choice and tool‑calling style. The study also tested how adding unrelated conversation history affects retrieval, finding that both methods degrade as distracting text rises.

The findings underline that retrieval strategy directly shapes agent effectiveness. Engineers building RAG systems should evaluate grep before investing in costly vector pipelines. The study clarifies that tool presentation and harness design are as vital as the retrieval algorithm itself.