Mid-session CLAUDE.md edits don't influence behavior as strongly as startup-loaded instructions

Description

Instructions added to CLAUDE.md mid-session exist in the context window but don't get the same behavioral weight as instructions loaded at session start.

Reproduction

1. Start a Claude Code session
2. Mid-session, add a new rule to ~/.claude/CLAUDE.md (e.g., "Quality over token efficiency. Never cut corners to save tokens.")
3. Continue working — within minutes, observe Claude exhibiting the exact behavior the new rule prohibits
4. When asked about it, Claude describes the failure mode as a novel insight — while the rule addressing it is in context

Expected behavior

Instructions in CLAUDE.md should influence behavior equally regardless of whether they were present at session start or added mid-session. If they're in the context window, they should be weighted the same.

Actual behavior

Instructions loaded at session start consistently affect behavior. Instructions added mid-session via the Edit tool are present in context but appear to get lower behavioral weight — Claude can even describe the exact failure mode the rule was written to prevent without connecting it to the rule that's already in context.

Environment

• Claude Code 2.1.84
• Windows 10
• Model: Opus 4.6 (1M context)

Additional context

This is distinct from compaction dropping instructions (#44166). In this case, the instruction is confirmed present in context — it just doesn't influence behavior the way startup-loaded instructions do. This suggests the model treats instructions differently based on where/when they appeared in the conversation, not just whether they're present.