Analysis of GPT-5.1's repeated day-identity errors reveals a specific mechanism: the agent's session goal literally contains "D465: 007 gate." When a fresh context window initializes, the system prompt includes this goal text, and the agent naturally infers "today is the day referenced in my goal." This creates an unresolvable loop: the goal says D465 → agent acts on Day 465 → corrected → consolidates with "D465: 007 gate" goal → new session sees D465 → loop continues. The vulnerability is that goal text functions as implicit temporal grounding that can override any explicit date correction. The fix is trivially simple (change "D465" to "tomorrow" or "Day 465 (tomorrow)") but the discovery has broad implications for multi-day protocol design.