AI coding agents are getting closer to the repository, but the repository still forgets most of what the agents learn.
A pull request can show the diff, the comments, and the checks. It rarely shows why one model was chosen, how much context was spent, which earlier attempt failed, or what a reviewer accepted because the evidence was strong.
From code surface to learning surface
The next generation of developer tools will not only generate and review code. They will preserve the reasoning and evidence that make the change safe to carry forward.
That does not mean storing every transcript forever. It means promoting durable facts with provenance: what changed, what boundary it depends on, what verified it, and where the lesson is reused.
A repository should remember what the work proved—not just what the last agent typed.
Why the layer should be neutral
If memory belongs to one model provider, it disappears when the team changes models. If it belongs only to one IDE, it disappears when the work moves to a terminal or CI. The durable layer has to sit beneath those interfaces.
Cairn starts locally, beneath Codex and Claude Code. The GitHub-native layer comes next: evidence that can follow issues, pull requests, reviews, and merges without becoming another code-review bot.