Causality-Graph Memory Outperforms Similarity-Based Retrieval for Long-Horizon LLM Agents
Existing memory benchmarks for LLMs are dialogue-centric and fail to represent the state-action-observation trajectories characteristic of real agentic workloads. AMA-Bench fills this gap by combining real-world agent trajectories with expert-curated QA and synthetic trajectories that scale to arbitrary horizons. The authors show that current memory systems underperform primarily due to lossy similarity-based retrieval and an inability to capture causal and objective information. Their proposed AMA-Agent, which uses causality-graph construction and tool-augmented retrieval, achieves 57.22% accuracy — an 11.16 percentage point improvement over the strongest baseline.