Heavy-Tailed Memory Traces in Long-Horizon Language Agents

Summary

Long-horizon language agents often use external memory as a frozen world model, but common evaluations focus on task success or token cost rather than how memory use is distributed. This study conducts a conservative tail audit under finite context and repeated retrieval, finding reproducible but policy-dependent concentration: random-walk agents produce log-normal-compatible retrieval artifacts, while semantic LLM policies show the strongest truncated-power-law-compatible core-tail traces. The authors introduce Core-Tail World Model (CTWM), a rank-based memory controller that uses a single exponent to allocate prompt budget while retaining a summarized tail. On Synthetic Graph World, CTWM preserves full state and transition coverage, cuts prompt tokens by 5.9%, and reduces bottom-half tail prediction error by 13.6% versus a graph-memory baseline. It also produces consistent token savings on ALFWorld and reduces tokens by 24.48% on LongMemEval while maintaining aggregate accuracy parity. The results position heavy-tailed memory traces as both a diagnostic of finite retrieval and a control signal for more token-efficient agent world models.