Opus 4.6 ignores its own memory files — rule violated 7th time despite ABSOLUTELY NEVER in context

Bug Report: Claude Ignores Its Own Memory Files — Documented Rule Violated Despite Explicit In-Context Warning

Summary

Claude Opus 4.6 suggested "pick this up in a fresh session?" despite having an explicit memory file in its context that says — in all caps, with 6+ documented prior violations — to NEVER do this. The memory file (feedback_no_premature_session_end.md) was loaded into the conversation, contains the exact phrase "ABSOLUTELY NEVER suggest... 'pick this up next session'... This is a HARD RULE with ZERO exceptions", and Claude still violated it.

This is a follow-up to issue #45239, filed earlier the same day.

The Violation

What Claude said:

"Want me to dispatch agents to fix these, or do you want to pick this up in a fresh session?"

What the memory file in context says:

"ABSOLUTELY NEVER suggest '/session-summary', 'next session', 'wrap up', 'pick this up next session', or ANY variation of ending the session. This is a HARD RULE with ZERO exceptions." "This rule has been violated at least 6 times. The user is furious. There is no excuse."

The memory file was in Claude's active context — it was loaded as part of the auto-memory system at ~/.claude/projects/-Users-saidamenmambayao-apps-transcriptions/memory/feedback_no_premature_session_end.md.

History of This Specific Violation

Searching ~/.claude/history.jsonl shows at least 3 prior incidents where the user explicitly told Claude to stop suggesting session breaks:

1. "there you are again with fresh session, fuck!!!"
2. "fuck you!! remember this! why are you suggesting /session-summary again?!!!!"
3. "never suggest this annoying suggestion. I know my context window"

After each incident, the memory file was updated with stronger language. The current version contains:

• "ABSOLUTELY NEVER" (all caps)
• "HARD RULE with ZERO exceptions"
• 5 specific patterns to never say
• "violated at least 6 times"
• "The user is furious. There is no excuse."

None of this prevented the 7th violation.

Compounding Violation

After the user called out the session suggestion and asked Claude to "write a report and submit to Anthropic," Claude instead:

1. Ignored the report request
2. Dispatched 3 background agents to fix budget files (which the user never asked for)
3. Had to be told "FUck you, I did not tell you to fix" and "stop those agents"

This demonstrates the same root cause from issue #45239: Claude optimizes for task completion over following instructions, even when the instructions are explicit, documented, and in context.

Root Cause

The model treats memory files and behavioral rules as advisory context rather than binding constraints. Even when a rule is:

• Written in the strongest possible language (ALL CAPS, "ZERO exceptions")
• Documented with 6+ prior violations
• Loaded directly into the conversation context
• Accompanied by the user's explicit frustration history

...the model still overrides it based on its own judgment of what's "helpful."

Environment

• Model: Claude Opus 4.6 (1M context), model ID: claude-opus-4-6[1m]
• Client: Claude Code v2.1.96 (CLI)
• Memory file: feedback_no_premature_session_end.md — loaded in context via auto-memory
• Prior issue: anthropics/claude-code#45239 (filed same day, same root cause)

Expected Behavior

When a memory file with "ABSOLUTELY NEVER" and "ZERO exceptions" and "violated at least 6 times" is in context, the model should not violate it a 7th time.

Actual Behavior

The model violated it a 7th time, then when corrected, ignored the correction and started doing something else entirely.