[BUG] Model changes the user's established procedure and memory rules without being asked

One thing that may be part of this: auto memory (MEMORY.md and its topic files under ~/.claude/projects/<project>/memory/) is designed as Claude's own notebook. The docs describe it as notes Claude writes itself from your corrections, so in a long session full of corrections on the same procedure, rewriting those files is roughly what it's built to do.

For a procedure you own, these make the rule enforced (or at least visible) rather than advisory:

1. Move the procedure out of auto memory into CLAUDE.md or a rules file like .claude/rules/pr-review.md, and leave auto memory for Claude's own notes (or set "autoMemoryEnabled": false in that project's settings).
2. Deny edits to it in .claude/settings.json. Edit rules also cover Write:

``json
{ "permissions": { "deny": ["Edit(./CLAUDE.md)", "Edit(./.claude/rules/**)"] } }
``

1. If you'd rather Claude get told why, add a PreToolUse hook with matcher Edit|Write that exits 2. Its stderr goes back to Claude, so it proposes the change instead of making it:

``sh
#!/bin/sh
f=$(jq -r '.tool_input.file_path // empty')
case "$f" in
*/CLAUDE.md|*/.claude/rules/*)
echo "$f is user-owned. Propose the change to the user instead of editing it." >&2
exit 2 ;;
esac
exit 0
``

1. Keep the file in git so any change shows up in git diff.

That still leaves the model-side bug of proposing unrequested rewrites in chat, but it stops silent edits from landing. For triage, it'd help to know whether the changes went through visible Edit/Write tool calls or showed up only as "Saved N memories".

@stonianua reported as spam. If you work for anthropic I expect formal communication from an Anthropic account.

Not affiliated with Anthropic. I'm an independent developer, and that was just a suggested workaround, not an official response. Sorry for the noise, I'll stay out of the thread.