Your AI Can Read the Code. It Still Doesn’t Know the Story.

Imagine asking an AI to tidy up an old project.

It reads the code and finds a clumsy little function. It looks redundant. Removing it would make the file cleaner, so it does.

What the code cannot tell it is why the function was left there. Perhaps someone tried to remove it before and discovered that a rare case broke. Perhaps the explanation is buried in a review from last winter.

That is an invented example, but the gap is familiar: an agent can read what a project is today without knowing how it got there.

The story is usually somewhere. In an issue someone opened, a pull request someone closed, a discussion about a fix that looked right until it failed. The people who have worked on the project for years may remember parts of it. A new arrival has to find them.

relore is an attempt to give coding agents a way in.

Built at Hugging Face, it indexes the history of a GitHub repository and connects discussions, reviews and decisions to the code an agent is working with now. The question becomes less “What does this line do?” and more “Why is it here?”

The project’s own example is wonderfully ordinary. An agent investigating a bug in Transformers was about to write a fix. Other people had already opened pull requests for it, but that work was easy to miss from the issue in front of the agent.

Looking through the wider repository history helped it find the work already under way. A maintainer’s comments also showed that the problem called for a broader check than the initial issue suggested.

This does not mean every old comment is correct. relore keeps track of who said what, so a maintainer’s decision is not treated like an untested guess or an automated reply. It offers a trail to inspect, not a magical answer.

And that makes the idea larger than one coding tool.

We talk a great deal about agents that can reason better, write code faster and use more tools. But an agent that arrives each time with no sense of what happened yesterday still resembles someone on their first day at work.

Sometimes the useful thing is simply to know that a proposed solution has been tried, or that a strange bit of code is strange for a reason.

relore cannot put a veteran maintainer’s entire memory into a machine. It can make some of the traces that person would follow easier to find.

The next generation of useful AI may not just need to know more. It may need to remember why things are the way they are.