Problem

The Copilot CLI agent can autonomously store "repo-wide memories" — facts it learns during a session that persist and are loaded into every future session for all users in the repository. There is currently:

• No way to delete a memory once stored
• No admin controls to manage memories across a repository
• No user attribution — memories don't record who triggered them
• No user approval — the agent decides on its own what to memorize

This creates a serious problem: any user can unintentionally pollute the repo-wide memory with user-specific preferences, jokes, or offhand comments, and there is no mechanism to clean them up.

Real-World Example

In Azure/azure-sdk-for-net, a user jokingly told the agent to add a specific footer (🤖 m-nash-copilot) to PR comments. The agent:

1. Took this literally and stored it as a repo-wide memory: "All PR/issue comments written on behalf of the user must end with \n---\n🤖 m-nash-copilot"
2. This memory was then loaded into every other user's sessions across the entire repo
3. When the official Copilot reviewer bot later left a comment on an unrelated PR, it included the m-nash-copilot footer — visible here: Refine mpg-sdk-migration skill: local iteration, sub-skills, and PR workflow Azure/azure-sdk-for-net#56515 (comment)
4. Another user (live1206) had to ask "why did you add '🤖 m-nash-copilot' in the end of the comment? I am not m-nash"
5. This triggered a second contradictory memory to be stored, and neither can be deleted

The only "fix" available today is storing yet another memory to override the bad ones — which just adds more clutter.

Requested Features

P0 — Memory Management

• Ability for repo admins to delete individual memories

P1 — User Attribution & Audit

• Record which user stored each memory and when
• Provide an audit log of memory additions

P2 — Guardrails on Memory Storage

• Require user confirmation before the agent stores a repo-wide memory (e.g., "I'd like to remember this for all users in this repo. OK?")
• Distinguish between user-scoped preferences (only loaded for that user) and repo-scoped conventions (loaded for everyone)
• Allow users to set a repo-level policy to disable autonomous memory storage

Impact

This affects any repository with multiple Copilot CLI users. Without these controls:

• One user's offhand remark can change agent behavior for the entire team
• There is no way to audit or clean up what's been stored
• The only workaround (storing corrective memories) adds noise without removing the original problem
• Trust in the agent's behavior is undermined when it acts on stale or incorrect memories