[BUG] Memory/feedback system does not reliably change agent behavior across sessions

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

I asked Claude Code to "commit tout et push" (commit everything and push). This required 7 manual approvals from me for what should be a single atomic operation.

This is despite having two separate memory/feedback entries explicitly instructing Claude to chain everything without pausing:

• feedback_commit_push.md — "Quand l'utilisateur dit commit et push, tout chaîner sans pause"
• feedback_stop_asking.md — "Instruction claire = exécution complète, pas de confirmation à chaque micro-étape"

What Should Happen?

git add + git commit + git push chained in minimal tool calls, as instructed by the saved feedback.

Error Messages/Logs

Steps to Reproduce

The memory/feedback system creates an illusion of learning. Users spend time correcting behavior and confirming it's saved, but the same patterns repeat in subsequent sessions. The agent even acknowledges this when
confronted — "oui, probablement" — but has no mechanism to actually enforce stored feedback.

This creates a frustrating loop:

1. User corrects behavior
2. Agent saves feedback to memory
3. Next session, agent ignores the feedback
4. User corrects again
5. Agent finds existing feedback, adds a duplicate
6. Repeat

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

Windows 11

Claude Code Version

Claude Opus 4.6

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

VS Code integrated terminal

Additional Information

_No response_