GitHub intelligence makes context a product surface

A coding agent that forgets the team's conventions starts every task with a tax: rediscover the test command, infer the naming style, and repeat the same repository history. GitHub's new intelligence layer treats that context as a product surface rather than an instruction pasted into every chat.

GitHub describes three pieces: Memory carries preferences and repository conventions across sessions, Dreaming proposes ways to keep context current, and code search reaches beyond the current checkout. The company says Copilot Memory is generally available on paid plans while Dreaming is opt-in preview.

Remembering is a permission decision

A memory can make an agent faster and more consistent, but it can also preserve a stale preference or expose context to a session that should not have it. GitHub says repository customization proposes changes to checked-in skills and instructions through pull requests. That review boundary is more important than the word memory.

The valuable context is not the context an agent can store. It is the context a team can inspect and correct.

Builders should decide what belongs in a durable instruction, what belongs in a repository file, and what should expire after a task. Track stale memories, make provenance visible, and measure whether context reduces rework rather than merely increasing the amount an agent remembers.