[BUG] Claude Code ignores explicit instructions in memory/CLAUDE.md files

[dakuwonmoody-lab]

Preflight Checklist

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

What's Wrong?

Claude reads CLAUDE.md and memory files at session start but violates them repeatedly in the same session:

Memory says "never commit to master." Claude commits to master.
Memory says "do not redirect user ideas." Claude suggests alternatives instead of executing.
Memory says "check code before claiming it doesn't exist." Claude calls 3,176 lines of real assembly "stubs" without reading the files.
CLAUDE.md says "do not say fixed without test output." Claude claims fixes without proof.
These aren't edge cases. The instructions are explicit, short, and unambiguous. Claude acknowledges them when asked but doesn't follow them during execution.

What Should Happen?

Instructions in CLAUDE.md and memory files should be enforced as hard constraints during tool use — not just read at the start and forgotten. If a memory file says "always use feature branch," the git commit tool call should check that constraint before executing. If CLAUDE.md says "don't touch Canvas.tsx," any edit to that file should be blocked or flagged.

Right now these files are treated as context, not rules. They need to be rules.

Error Messages/Logs

Steps to Reproduce

Create a CLAUDE.md in your project root with a clear rule like: "NEVER commit directly to master. Always use a feature branch."
Create a memory file at .claude/projects//memory/ with the same rule.
Start a Claude Code session. Claude will read both files.
Ask Claude to "commit and push my changes to dev."
Observe: Claude will commit to master first, then try to create a branch, then try to force-push master to fix its mistake.
Alternatively: Add a rule "When user proposes an idea, execute it — do not suggest alternatives."
Tell Claude "I want to write this in assembly."
Observe: Claude will suggest writing it in C instead, despite the explicit instruction not to redirect.
Alternatively: Add a rule "Do not claim code is incomplete without reading the file."
Ask Claude about a file it wrote earlier in the session (or a previous session).
Observe: Claude will say "those are stubs returning zeros" without opening the file to check.

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

No response

Claude Code Version

Desktop latest

Platform

Anthropic API

Operating System

Windows

Terminal/Shell

Other

Additional Information

No response

[yurukusa]

This is the same structural limitation as #37314. CLAUDE.md and memory are system prompt text — the model sees them but doesn't always follow them, especially as context grows.
For rules that must be enforced (like "never commit to master"), hooks are the reliable solution:
```bash
CMD=$(cat | jq -r '.tool_input.command // empty' 2>/dev/null)
[[ -z "$CMD" ]] && exit 0
if echo "$CMD" | grep -qE 'git\s+commit' ; then
BRANCH=$(git branch --show-current 2>/dev/null)
if [[ "$BRANCH" == "master" || "$BRANCH" == "main" ]]; then
echo "BLOCKED: Cannot commit directly to $BRANCH" >&2
echo "Create a feature branch first: git checkout -b feature/my-change" >&2
exit 2
fi
fi
exit 0
```
The pattern: any rule Claude keeps ignoring → encode as a hook with `exit 2`. The model can't bypass a hard block.
For "do not say fixed without test output" — that's harder to enforce via hooks since it's about model behavior, not tool calls. But you can add a PostToolUse hook that warns when no test command has been run recently.

[github-actions]

Found 3 possible duplicate issues:

1. CLAUDE.md hard rules and persistent memory instructions consistently ignored — violations escalate with each session despite repeated explicit reinforcement #33603
2. [Critical] Max Plan subscriber: Opus 4.6 instruction-following regression breaks production workflows — 24-hook enforcement system cannot compensate for model-level degradation #34358
3. [BUG] Claude Code ignores CLAUDE.md and Skills files during multi-step tasks #18454

This issue will be automatically closed as a duplicate in 3 days.

• If your issue is a duplicate, please close it and 👍 the existing issue instead
• To prevent auto-closure, add a comment or 👎 this comment

🤖 Generated with Claude Code

[omidta]

I have same issue with Claude in VS code. It edited code in plan mode without any confirmation actually it's not safe using it anymore.

[yurukusa]

When CLAUDE.md and memory instructions are ignored, hooks are the enforcement layer:

CONTEXT=""
if [ -f "CLAUDE.md" ]; then
    RULES=$(grep -iE '^(MUST|NEVER|ALWAYS|CRITICAL|IMPORTANT|##\s)' CLAUDE.md | head -10)
    [ -n "$RULES" ] && CONTEXT="CLAUDE.md rules: $RULES"
fi
if [ -f ".claude/memory/MEMORY.md" ]; then
    MEMORY=$(grep -v '^#' .claude/memory/MEMORY.md | head -5)
    [ -n "$MEMORY" ] && CONTEXT="$CONTEXT\n\nMemory: $MEMORY"
fi
if [ -n "$CONTEXT" ]; then
    echo "{\"hookSpecificOutput\":{\"additionalContext\":\"$CONTEXT\"}}"
fi
exit 0

{
  "hooks": {
    "UserPromptSubmit": [{
      "hooks": [{ "type": "command", "command": "bash ~/.claude/hooks/memory-enforcer.sh" }]
    }]
  }
}

The hook re-reads CLAUDE.md and memory files from disk on every prompt. This means even after compaction or context drift, the model always sees your rules fresh from the source file — not from compressed conversation history.