On 28 October 2026 GitHub announced GitHub intelligence, which it describes as a “context and learning layer” for its AI coding agents. It has three parts, and they matter to anyone paying for software built with AI.
Memory (generally available, paid Copilot plans). Copilot now remembers personal preferences and repository conventions across sessions, whether it is used in the Copilot app, the command line, VS Code, code review, Slack, Teams or GitHub Mobile. GitHub’s own figures from the public preview: in reviewed command-line sessions where the agent used memory, 93% showed better adherence to developer preferences and 42% showed more efficient workflows. Those are GitHub’s numbers, not an independent test.
Dreaming (public preview, opt-in). Copilot reflects on past work and updates what it remembers. For team-level guidance, such as checked-in skills and instruction files, it proposes changes as pull requests for people to review. It can also turn recurring accepted code-review feedback into guidance. Dream runs consume AI credits, so there is a running cost.
Code search. Copilot can search code, configuration, schemas and documentation across related repositories without a local copy, including by what the code does rather than by exact wording.
What it means
The first generation of AI coding tools started every session from zero. This one accumulates habits. That makes output more consistent, but it also means a supplier’s AI setup becomes an asset that builds up over time, and something that can quietly drift if nobody reviews it.
So what
Ask any team building your product with AI tools two questions: what does the tool remember about our codebase, and who reviews changes to those instructions? Good answers mean faster, more consistent delivery. No answer means your project is being shaped by settings nobody is watching. Our AI-assisted development service is built around reviewed, owned AI workflows. To discuss yours, start a project.