What they did

The authors developed a system-oriented taxonomy to classify agent memory systems along four axes. They built a phase-aware profiling harness that attributes costs to construction, retrieval, and generation phases. They characterized ten representative memory systems (spanning flat retrieval, LLM-mediated extraction, consolidating fact stores, and agentic control flows) using two benchmark suites designed for long-horizon tasks. Experiments measured latency, memory overhead, and throughput.

Key findings

• The profiling revealed that design choices significantly shift cost between write (construction) and read (retrieval) paths; some systems optimize for fast retrieval at the expense of expensive construction, and vice versa.
• A key tradeoff exists between memory freshness and retrieval latency; systems that maintain fresh memory often incur higher construction latency.
• Query volume amortizes construction cost; systems with higher fixed construction overhead become more efficient under high query rates.
• The study derives 10 concrete system recommendations, including construction scheduling to avoid peak load, establishing capability floors for minimal acceptable accuracy, and fleet-scale management strategies.

Why it matters

This work provides the first empirical foundation for engineering memory systems in long-horizon agent deployments. By characterizing the system-level implications of memory design choices, it enables more informed decisions when building scalable agent infrastructure. The recommendations offer practical guidance for optimizing memory in production systems.

Caveats

The characterization is limited to ten systems and two benchmark suites, which may not cover all existing or emerging memory architectures. The benchmarks focus on specific long-horizon tasks; generalizability to diverse real-world workloads is not fully established. The profiling harness itself may introduce overhead not representative of all deployments.