Trajectory-Based Drift Detection in Stratified Agent Memory: A Minimal Deterministic Note on Coherence Decay Across Time

This preprint documents a minimal method for detecting internal drift in stratified agent systems by treating memory updates as a trajectory over time rather than evaluating them only at isolated steps. The framework operates within a three-tier memory structure in which short-term volatility, intermediate accumulation, and longer-horizon structural state evolve under asymmetric update rates. It frames coherence not as per-step admissibility, but as continuity across the trajectory of these coupled updates. Drift is defined as measurable deviation in that trajectory, particularly when recent accumulation begins to separate from structural lag or when update direction breaks relative to its prior course. A deterministic micro-run is used to show that contradiction produces a staged pattern—directional instability at onset, delayed cross-tier misalignment, and subsequent reversal during recovery—without introducing stochastic behavior or additional subsystems. The purpose is to make time-extended coherence decay visible as an inspectable signal grounded in existing stratified memory dynamics. This work serves as a compact extension to constraint-governed agent architectures and is published as a technical preprint. Version v0.2 — deterministic demonstration release.