AI coding agents are getting memory.
That sounds small, but it is one of the biggest practical issues for businesses using AI in software work.
GitHub updated Copilot Memory with better deletion guidance, a repository-level off switch, /memory on, /memory off, and /memory show in Copilot CLI, plus clearer prompts showing whether a memory is personal or repository-level.
My take: memory is powerful, but unmanaged memory is a liability.
If an AI agent remembers project rules, coding style, deployment steps, or business context, it can save real time. But teams also need control:
- Who can see this memory?
- Can it be deleted?
- Is it personal or shared with the repo?
- Can repository owners turn it off?
- Does it store anything sensitive?
This is where AI agents are moving: not just “write code for me,” but persistent workflows that understand a project over time.
The practical checklist
Before enabling memory in an AI coding agent, use this checklist.
1. Separate personal preferences from repository facts
Personal memory should be things like: “I prefer concise explanations.”
Repository memory should be things like: “This project uses Laravel, PostgreSQL, and GitHub Actions.”
Do not mix private business details into shared project memory.
2. Decide what memory is allowed
Allowed:
- coding standards
- test commands
- deployment checklist
- folder structure
- safe project conventions
Not allowed:
- passwords
- API keys
- customer data
- private contracts
- payment details
- anything copied from a confidential email
3. Make deletion obvious
If an agent remembers the wrong thing, the team must know where to delete it.
If nobody knows how to remove a memory, the feature is not ready for production use.
4. Add a repository-level off switch
Some projects should not use memory at all.
Examples: client projects, regulated work, sensitive internal systems, payment flows, or anything with personal data.
5. Review memory like documentation
Memory is not magic. Treat it like project documentation that can become outdated.
Review it when the stack changes, the deployment process changes, or the business rules change.
6. Test agent behavior
Do not ask “does it feel better?” Ask measurable questions:
- Did the agent follow the project rules?
- Did it run the expected tests?
- Did it avoid secrets?
- Did it ask for approval before risky changes?
- Did it leave clean output?
The big idea: AI agents are not just chatbots anymore. They are becoming persistent project workers.
That means businesses need the boring stuff: scope, permissions, deletion, review, testing, and human approval.
Boring is good. Boring keeps systems from catching fire.
Sources
- GitHub Changelog: https://github.blog/changelog/2026-05-26-copilot-memory-has-more-controls-for-deletion-scope-and-the-copilot-cli/
- GitHub Copilot Memory docs: https://docs.github.com/copilot/concepts/agents/copilot-memory
- OpenAI Developers — Testing Agent Skills Systematically with Evals: https://developers.openai.com/blog/eval-skills
- OpenAI Developers — Using skills to accelerate OSS maintenance: https://developers.openai.com/blog/skills-agents-sdk