Two kinds of memory: the repo’s and yours
GitHub launched Copilot Memory (first called agentic memory) in public preview on January 15, 2026. As of November 2026 the docs describe two scopes:
| Scope | What it holds | Who benefits | Where it applies |
|---|---|---|---|
| Repository facts | Coding conventions, build and test commands, architectural decisions | Everyone with access to that repo and Copilot Memory | Only in that repository; forks do not copy them |
| User preferences | How you like answers formatted, which actions need your approval, workflows you reuse | Only you | Across all your repositories |
Copilot creates memories on its own while it works. You can also state a fact or preference, or simply ask it to remember something. Writing a repository fact requires write access to the repo.
Each memory carries a citation: a repository fact points to the code that supports it, a user preference points to your own words. Before Copilot relies on a memory, it checks that citation against the current branch. Memories that hold up get upvoted; wrong or outdated ones get downvoted and drop out of the list.
Where Copilot uses it
Repository facts work in the GitHub Copilot app, Copilot CLI, Copilot code review, the Copilot cloud agent, Agentic Autofix, Copilot in GitHub Mobile, Slack and Microsoft Teams. User preferences work in the same places except code review and Agentic Autofix. A memory captured in one place can be used in another: something the cloud agent learned can show up in your CLI session.
In VS Code it runs through the setting chat.copilot.memory.enabled. VS Code also has its own local memory files (/memories/, /memories/repo/, /memories/session/), which live on your machine and are not shared. Those belong to the older Local harness, which VS Code plans to remove in February 2027.
How long memories last
By default a memory expires after 90 days unless it is reinforced. Validation and upvoting reset the clock; simply being retrieved does not. At launch in January the window was 28 days, so older blog posts quote that number. Enterprise and organization owners can set the expiry for repository facts anywhere from 7 to 180 days; user preferences always use the default.
There is also Copilot Dreaming, in public preview: a background process that reviews sessions and pull request feedback, merges duplicates and removes stale memories without asking for approval per change. Changes it suggests to checked-in instructions arrive as a pull request. It is not available on Copilot Pro.
How to turn it on, off, or delete memories
Who controls it depends on your plan:
- Copilot Pro, Pro+ and Max. User preferences are on by default. Change it under your profile picture > Copilot settings > Features > Copilot Memory, with the options User memories allowed, Repository memories allowed, User and repository memories allowed or Don’t use memories.
- Copilot Business and Enterprise. User preferences stay off until an organization or enterprise owner enables the policy. Repository facts are allowed by default unless the enterprise, organization or repository restricts them.
- Delete your own preferences. Copilot settings > Memory, then the trash icon or select several and click Delete.
- Delete repository facts. Repository Settings > Copilot > Memory. This requires write access. Repository admins can also switch memory off for that repo there.
- Copilot CLI.
/memory offturns both types off,/memory onturns them back on,/memory showshows the setting.
Switching memory off does not delete what is stored. Admins on Business and Enterprise can view, export and delete members’ preferences.
The common misunderstanding: memory replaces instructions
Many developers expect Copilot Memory to make .github/copilot-instructions.md unnecessary. GitHub says the opposite: Copilot may not apply a memory in every session, so rules that must always hold belong in custom instructions. Memory is for what Copilot picks up along the way; instructions are what the team agrees on and reviews. If you want to know how the community tries to bridge that gap with a folder of notes, see Do you need a memory bank for GitHub Copilot?.
What this means for a team
Copilot Memory is the first Copilot feature that shares something learned between colleagues: a repository fact found in one person’s session helps the next person in that repo. But it stops at the repository boundary. What Copilot learned in your backend repo does not reach the frontend repo, and user preferences stay personal. It also stays inside GitHub Copilot. When half the team works in Claude Code or Cursor, those tools start from zero, even in the same repo. For a setup that keeps the shared rules in one file across tools, see How to share AI context in a dev team.
Where Bombos fits
Copilot Memory remembers per repository and per tool. Bombos is a plugin that gives Copilot and the other AI tools your team uses (Claude Code, Codex, Cursor, ChatGPT, your own agents) one shared memory: learned rules and team knowledge that carry across repositories, plus what already sits in documents and mail, traceable, in the cloud or on your own server. One developer can try it first with their own agents. More on the idea in Shared memory across coding agents.
Related questions
Does Copilot Memory work in VS Code?
Yes, through the setting chat.copilot.memory.enabled. Separately, VS Code has local memory files under /memories/ that stay on your machine and are not shared with teammates.
How do I turn off memory in Copilot CLI?
Type /memory off in a session, or set "memory": false in your Copilot CLI user settings. For non-interactive runs with copilot -p, memory is off unless you add --enable-memory.
Can I write or edit a memory myself on GitHub.com?
No. You can ask Copilot to remember something during a session, but on GitHub.com you can only view and delete memories, not create or edit them.
Does Copilot Memory store secrets?
GitHub rejects memories that contain detected secrets, and Copilot is instructed not to store sensitive personal or confidential information or short-lived task instructions.
Do memories cost anything?
There is no separate fee, but memories count toward token usage and AI credit consumption, like any other context Copilot reads.
