IS GREP ALL YOU NEED? THE HARNESS MATTERS MORE THAN THE SEARCH

Summary from the TLDR Dev newsletter

Lexical search using grep frequently outperforms vector-based semantic search in long-memory question answering tasks. The agent harness, or the framework used to deliver results to the model, is just as influential as the retrieval method itself. Exact string matching is great in these scenarios because it captures verbatim details like names and dates that semantic embeddings often miss or smooth over.