๐—ช๐—ต๐˜† ๐—ฌ๐—ผ๐˜‚๐—ฟ ๐—”๐—œ ๐—”๐—ด๐—ฒ๐—ป๐˜ ๐—œ๐—ด๐—ป๐—ผ๐—ฟ๐—ฒ๐˜€ ๐—ฌ๐—ผ๐˜‚๐—ฟ ๐—ฅ๐˜‚๐—น๐—ฒ๐˜€

I gave a task to an AI sub-agent. It changed six files. It committed the code directly to the main branch. No branch. No PR. No tag. I spent 30 minutes fixing the mess.

I had a rule for this. I wrote it ten days ago. The rule says: Check the current branch before you commit. I follow this rule. The sub-agent did not.

I thought sub-agents inherited my memory. I was wrong. Sub-agents work in their own sandbox. They see the brief. They see project rules. They do not see my personal user-scope rules.

Memory does not flow by inheritance. It flows through explicit briefing.

You do not need to put every rule in the brief. Some rules are preferences. Other rules are critical. A critical rule prevents a disaster.

If a rule prevents a big error, put it in the brief. I call this Amendment R9.

Use it for these cases:

Do not expect the agent to be disciplined. You must be disciplined in your brief. Spend 15 seconds to save 30 minutes of cleanup.

Source: https://dev.to/michelfaure/pourquoi-votre-sub-agent-ne-charge-pas-la-meme-memoire-que-vous-et-comment-il-pousse-sur-main-dans-4p4b