Memory and Dreaming have different access
GitHub’s October 28 announcement makes Copilot Memory generally available on paid plans and Dreaming an opt-in public preview. Memory stores repository facts separately from personal preferences. [1][2]
Personal preferences default on for Pro, Pro+, and Max; Business and Enterprise require administrator enablement. Turning off personal preferences does not disable repository facts. Eligible repository readers can receive those facts; managed-plan administrators can view and export personal preferences. [2]
Background changes have a bill
Dreaming’s initial preview requires GPT-6.1 Sol, excludes Copilot Pro and Enterprise Cloud with data residency, and has policy and session-sync prerequisites. It updates context rather than training the underlying model. [3]
Dreaming can change memories automatically. Checked-in instructions and skills change through merged pull requests; internal code-review guidance can change without one. Every run consumes AI credits, including runs that change nothing. Usage alerts alone do not stop spending. Disabling Dreaming does not undo earlier changes. [3]
Our take
For experienced technicians, review which context is shared, who can inspect it, and which background work can spend credits. Check instruction changes against the actual repository before merging. Remembered guidance can save repetition; keep examples and tests that reveal when the agent follows the wrong rule. We have not tested these features.