Contents

GitHub Copilot Dreaming aims to keep AI instructions useful as projects evolve by reviewing past work and suggesting updates in the background. According to Publickey’s November 5 report, GitHub announced the feature in public preview at GitHub Universe 2026.

Key Points#

  • Dreaming examines Copilot conversations and pull request reviews to identify instructions and memories that need updating.
  • These include repository coding conventions, architectural knowledge, and project-specific commands and rules.
  • In the keynote demo, Dreaming proposed a pull request to replace Java instructions after an application migrated to Svelte and JavaScript.
  • Check GitHub’s documentation for preview eligibility, settings, data management, and approval requirements.

What Dreaming Updates#

Instructions can become outdated even when nobody edits them. A repository may move to another framework, build commands may change, or a team may adopt new coding conventions, yet Copilot continues receiving instructions tailored to the previous setup.

Dreaming addresses this maintenance problem. According to the report, it covers AI instructions and information stored in Copilot Memory, including repository architecture, coding rules, and project-specific commands. GitHub also cites changes in model capabilities as a reason to revisit existing instructions.

The key distinction is between retaining information and keeping it accurate. Remembering an old command that a project no longer supports is not particularly useful. Dreaming is intended to revisit stored context as work progresses.

How It Learns from Work#

According to the report, Dreaming analyzes interactions with Copilot and pull request reviews, using that information to organize and update instructions. This allows it to draw on both code changes and instructions developers repeatedly give during their work.

For example, repeated corrections in reviews may indicate that existing instructions no longer match the team’s expectations. This is a plausible use of the reported mechanism, but the article does not explain how Dreaming handles conflicting feedback or decides which comments should become lasting rules.

Background processing may reduce the effort of finding outdated instructions. Even so, reviewers still face an important question: does a proposed change reflect an ongoing project requirement or a decision specific to one task?

A Migration Example#

The keynote demo featured an application rewritten from Java to Svelte and JavaScript. Its instruction file still assumed it was a Java project. Dreaming identified this mismatch and proposed a pull request to rewrite the instructions to match the new implementation.

This example makes the benefit concrete. Migrating code creates another maintenance task: updating the context that tells an AI assistant how to work with it. If overlooked, assumptions the repository no longer uses may carry over into subsequent requests.

Pull requests also give developers a way to inspect the changes. In this demo, the proposed instructions appeared as changes available for review. However, the report does not clarify whether every type of Dreaming update follows the same approval process.

What to Check in the Preview#

As this is a public preview, it is best to keep the evaluation limited in scope. The article explains its purpose and one workflow, but lacks enough detail to assess operational controls or availability for a particular account.

Before enabling it, check these four points in GitHub’s latest documentation.

  • Confirm which plans, accounts, and repositories are eligible for the preview.
  • Understand which conversations, reviews, and repository information are analyzed.
  • Check where update proposals appear and which changes require approval.
  • Check how to correct, disable, or revert unwanted updates.

For an initial evaluation, choose one repository with known outdated instructions, such as retired build commands or references to an old framework left after a migration. Before testing, record the current instructions and the correct replacements. Then assess whether Dreaming finds the mismatches and proposes accurate, appropriately scoped fixes.

A suitable proposal must preserve rules that remain valid. For example, removing Java-specific instructions should not automatically remove unrelated security requirements. Review the entire change rather than accepting it simply because it recognizes the new framework.

Next Steps#

Choose one repository and record one outdated AI instruction alongside its correct replacement to create a test case for evaluating the Dreaming preview.

Source: https://www.publickey1.jp/blog/26/github_copilot_dreaminggithub_universe_2026.html

Contents