Two years ago, in a repository nobody has opened since, you ruled out SQLite for service state. The code records your eventual choice. The reasoning is gone, along with the alternatives you rejected on the way. Now an AI agent is working in another repository. It can't see that earlier decision, so it proposes SQLite again, with full confidence.

Agent memory tools don't solve this. They retrieve facts and preferences based on similar wording, but a decision is mostly what you didn't choose: the options you rejected, why, and what would have to change before you reconsidered. Similarity search can pull precedent from the wrong kind of project. If the same bad choice appears in eight repositories, it can mistake repetition for a recommendation and hand the agent that choice a ninth time.

Coding agents need access to earlier reasoning without blindly applying every past decision. In this session, you'll see what a useful decision record contains, how an agent works out which projects it applies to, and how deliberate exceptions keep old warnings from becoming permanent noise. I wrote Precedent, an open-source plugin for Claude Code, Codex, and Cursor, and I'll run the same agent with and without it.

For engineers who use AI coding agents across many codebases. No ML background needed.