Latest in AI
GitHub Copilot Memory and Dreaming: Availability, Costs and Controls
Copilot can retain and revise working context, but automatic memory changes and reviewed instruction changes follow different rules.
GitHub made Copilot Memory generally available on paid plans on October 28 and introduced Dreaming in public preview. Memory carries preferences and repository knowledge between sessions. Dreaming maintains that context in the background, with important restrictions on plans, model access and how changes are approved.
What GitHub intelligence includes
The GitHub intelligence announcement groups Memory, Dreaming and cross-repository code search into a context layer for agents. Memory can preserve conventions across Copilot clients; code search helps locate relevant implementations without first cloning every repository.
That addresses a familiar problem with coding agents: the reasoning may be capable, but the session starts without the team's accumulated knowledge. A correct test command or a note explaining a compatibility constraint can be more useful than another page of generic instructions.
Memory being GA does not mean every related capability is GA. Dreaming is optional and remains a preview.
Dreaming changes context, not model weights
The Dreaming documentation explicitly distinguishes background context maintenance from model training. It can inspect eligible work and review feedback, then curate personal memories, repository knowledge and code-review guidance.
Not every change waits for a pull request. Memories can be added, revised or removed automatically, as can internal review guidance. Changes to checked-in instructions and skills arrive as pull requests and do not take effect until merged.
This makes the review boundary important. Teams already inspect changes to an instruction file, but less visible context can also influence what an agent does next. A useful rollout should include periodic checks of remembered assumptions, not only approval of proposed files.
Availability and billing restrictions
The initial Dreaming preview excludes Copilot Pro and GitHub Enterprise Cloud with data residency. It requires GPT-6.1 Sol access and applicable policy permission. Personal curation and repository customization also require session syncing; code-review customization has different prerequisites and does not support self-hosted runners.
Runs consume AI credits even when they produce no changes or process inactive repositories. Disabling Dreaming stops future work within the relevant setting's scope, but does not reverse earlier changes. These are reasons to enable a small, useful set of repositories first rather than everything by default.
What should improve in practice
For a team maintaining several services, the promising outcome is fewer repeated corrections: the agent uses the right test command, respects a known migration constraint and retrieves the relevant neighboring implementation.
That should be measured through the quality of the next change, not the quantity of memories created. Persistent context is useful only while someone can tell whether it still reflects the system.
I'm Paras. I work at the intersection of AI, engineering and distribution.
I help technical teams turn product updates and engineering decisions into stories developers can understand and share.