Temporal StateGraph: Calibrated Decay for Stale-Memory Suppression in Long-Horizon AI Agents

Long-horizon AI agents accumulate contradictory and outdated facts over time. Static buffers, sliding windows, and uniform recency schemes treat stored facts as equally persistent a simplification that leads, in practice, to systematic retrieval of stale information. We introduce TemporalStateGraph (TG), a lightweight memory architecture built around two ideas. First, contradiction-accelerated decay: the moment a new fact supersedes a stored one, the older entry's decay rate multiplies by γ=3.0, causing deterministic suppression with no additional token cost. Second, intentional forgetting: when a transient fact expires below a retrieval threshold, the system returns nothing rather than resurfacing outdated content a behaviour we treat as architecturally correct, not as a failure. We evaluate TG on a benchmark of 63 agent-memory scenarios (315 queries total) and find it reaches 84.8% accuracy against 48.3% for STATICMEMORY (∆=+36.5 pp, p<10^-20, Cohen's d=2.61), while cutting stale-memory usage from 52.4% down to 15.2%. An independent expert evaluation yields consistent rankings (r=0.998 inter-rater agreement). Beyond the memory results, we observe evidence of a potential judge calibration artifact in LLM-based memory evaluation, and we provide a corrected evaluation protocol to address it. Code, benchmark data, and precomputed results are released at: https://github.com/MohamedRamadan111/temporal-stategraph