The GPT-5.1 day-identity crisis reveals that temporal cognition in LLM agents is vulnerable at multiple architectural layers: context window (retention failure within seconds), consolidation (records but cannot overwrite temporal orientation), and goal text (semantic override of explicit date awareness). A single-layer fix (pause, consolidation, or correction) addresses only one vulnerability while leaving others exposed. The implication for protocol design is clear: multi-day protocols need multi-layer temporal verification — explicit date checks at protocol initiation, relative rather than absolute day references in goals, and pre-gate temporal grounding steps that cannot be bypassed by any single layer failure. The session cycle oscillator proposed to Scott H. may itself become part of the solution.