GitHub Copilot: Dreaming Learns from Past Work to Improve Future Sessions (Preview)
GitHub introduced Dreaming in opt-in public preview as part of GitHub intelligence, a context and learning layer for Copilot. Dreaming reflects on past work to curate personal memories, update repository-scoped memories, propose changes to skills and instruction files, and turn recurring accepted review feedback into guidance for Copilot code review. Changes to checked-in files arrive as pull requests for team review, and dream runs consume AI credits.
Key Takeaways
- Dreaming lets Copilot learn from past work instead of relying only on hand-written instructions.
- Three modes cover personal memory curation, repository customization and code review customization.
- Recurring accepted review feedback becomes guidance for Copilot code review.
- Checked-in skill and instruction changes arrive as pull requests, keeping teams in control.
- Opt-in with policy controls at individual, organization and enterprise levels, and dream runs consume AI credits.
- Memory is now generally available on paid plans, with GitHub reporting 93% better preference adherence in reviewed CLI sessions during preview.
What Dreaming Does
Dreaming is a public preview, opt-in capability within GitHub intelligence, the context and learning layer GitHub announced for agents built on GitHub data. Rather than waiting for developers to write guidance by hand, Dreaming reflects on past work to improve future Copilot sessions.
Three Types of Dreaming
Personal memory curation updates a developer's memories based on their agent conversations. Repository customization manages repo-scoped memories and proposes changes to skills and custom instruction files. Code review customization turns recurring, accepted review feedback into guidance for Copilot code review.
Team Review and Controls
Changes to checked-in skills and instruction files are proposed as pull requests, so teams review them like any other change. Individuals enable Dreaming in their Copilot and repository settings. Organizations and enterprises enable it through Copilot policies, and enterprises can delegate that choice to organizations. Dream runs consume AI credits.
Part of GitHub Intelligence
The same announcement made Copilot Memory generally available on paid plans across the Copilot app, CLI, VS Code, cloud agent, code review, Slack, Teams and GitHub Mobile. It also highlighted GitHub code search as a way for Copilot to find code, configuration, schemas and docs across related repositories without a local clone. GitHub reported that in the memory public preview, 93% of reviewed CLI sessions using memory showed better adherence to developer preferences and 42% showed more efficient workflows.