AI
Heavy-Tailed Memory Traces in Long-Horizon Language Agents
Publication: cs.AI updates on arXiv.org
Author: Xinyuan Song, Zekun Cai
Source published: 02/10/2026, 04:00:00 UTC
Briefing created: 02/10/2026, 04:07:52 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 02/10/2026, 04:03:23 UTC.
Briefings
This article investigates how long-horizon language agents allocate memory under finite context, revealing that memory use concentrates on a core while accumulating errors in a long tail. The study uses a conservative tail audit to compare random-walk agents and semantic LLM policies, finding that semantic policies produce stronger truncated-power-law traces. The authors propose the Core-Tail World Model (CTWM), a rank-based controller that reduces prompt tokens by 5.9% on synthetic graphs and achieves 24.48% reduction on LongMemEval while maintaining accuracy parity.
Key points
- Semantic LLM policies produce the strongest truncated-power-law-compatible core-tail traces in memory allocation under finite context.
- The Core-Tail World Model (CTWM) reduces prompt tokens by 5.9% on synthetic graphs and lowers bottom-half tail prediction error by 13.6% compared to a graph-memory baseline.
Context and limitations
Results are based on synthetic graph and ALFWorld benchmarks; token savings on LongMemEval are paired comparison only; aggregate accuracy parity is confirmed but tail prediction error reduction is specific to CTWM's bottom-half performance.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 02/10/2026, 04:07:52 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeigipycqzhlabyvmlooaknm5v23b2itrutnovsycko3duyricf5sku