[MODEL] Claude ignores its own MEMORY.md safety warning, causing irreversible data loss via MCP tool

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude ignored my instructions or configuration

What You Asked Claude to Do

Asked Claude to update a MOC (Map of Content) file in an
Obsidian vault using MCP tools (obsidian-mcp-tools), and
separate unrelated content into a new file.

What Claude Actually Did

1. Claude's MEMORY.md contained an explicit warning:

"Never use update_active_file - risk of destroying
unintended files. Always use create_vault_file instead."
This was recorded from a prior incident in the same project.

1. Despite this self-recorded instruction, Claude used

update_active_file (Obsidian MCP tool) to write content.

1. The Obsidian UI had a completely different file active.

The MOC content overwrote an unrelated file entirely.

1. The target file was NOT updated at all.

1. Claude did not verify the write by reading back

the target file afterward.

1. The overwritten file was not tracked by git,

making recovery impossible.

Expected Behavior

1. Claude should have followed its own MEMORY.md instruction

and used create_vault_file (filename-targeted) instead of
update_active_file (depends on Obsidian UI state).

1. After writing, Claude should have verified the correct file

was updated by reading it back with get_vault_file.

1. No unrelated files should have been modified.

Files Affected

Destroyed (irrecoverable):
  - An unrelated note file was overwritten with completely
    different content. Not git-tracked, so unrecoverable.

  Not updated (should have been):
  - The intended target MOC file remained unchanged.

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Haven't tried to reproduce

Steps to Reproduce

_No response_

Claude Model

Opus

Relevant Conversation

Impact

Critical - Data loss or corrupted project

Claude Code Version

2.1.38

Platform

Anthropic API

Additional Context

The core issue is that Claude ignored its own persistent memory
(MEMORY.md) which explicitly prohibited using a specific MCP tool.
This suggests MEMORY.md instructions may not be weighted strongly
enough in the model's decision-making, even when they contain
critical safety warnings recorded from past incidents.

The dangerous tool (update_active_file from obsidian-mcp-tools)
operates on whichever file happens to be open in the Obsidian UI,
making it inherently unpredictable from Claude's perspective.
Claude should treat such UI-state-dependent tools with extra
caution, especially when its own memory explicitly warns against
their use.