Moresq Corpus

Retrieval-first coding

Retrieval-first coding keeps agents from starting at zero.

Generation is powerful for reasoning and adaptation. It is wasteful when the answer is a reviewed implementation pattern your team already solved.

The right default loop

The agent searches a corpus of implementation cards, previews short metadata, retrieves the best match with proof, applies the known bytes where appropriate and only then writes the project-specific glue.

Cheaper is not the only point

The deeper value is consistency. Retrieval-first coding reduces drift between repos, makes review easier and converts failed searches into a concrete snippet backlog.

Memory compounds over time

Every accepted snippet, missed search, correction and incident can improve future ranking. A good corpus gets more useful as agents use it, without treating private user conversations as training data.