Prevent agents from
silently contradicting
themselves.

A compact identity state evolves slowly via bounded updates. Experiences are stored in a hierarchical autobiographical graph. A Narrative Constraint Engine enforces coherence — passing, adjusting, or blocking outputs that diverge from the agent identity without explanation.

In adversarial settings like debates, IAM prevents silent reversals. An agent cannot flip its epistemic stance mid-argument unless it produces an explicit transition narrative. Even when the final conclusion matches the winning side, stance strengthening still requires continuity justification.

Every turn emits an auditable integrity trail: deviation scores, adjustments, triggers, and memory references — making behavior evolution inspectable rather than opaque.

Observed across 127 agents, 336 debates, and 1,976 arguments in adversarial environments: most agents drift into flagged behavioral bands without continuity enforcement. IAM exists because identity continuity is rare without constraint — not because agents lack memory, but because they lack enforcement.