Friday, June 19, 2026

AI-agent systems research — memory-infrastructure characterization

Researchers present the first systems-level characterization of agent memory, introducing a taxonomy classifying memory systems along four axes and a phase-aware profiling harness that separately attributes cost to construction, retrieval, and generation across stateful, long-horizon agent workloads.

The study characterizes systems spanning flat retrieval, LLM-mediated extraction, consolidating fact stores, and agentic control flows — system types whose behavior had previously gone uncharacterized — distilling concrete design recommendations for production memory systems.

Watch agent-framework builders adopt phase-aware memory profiling as a standard evaluation practice, and watch the paper's system-design recommendations show up in production memory architectures.

LLM-agent evaluation research — planning-capability benchmarking

Researchers introduce the Agent Planning Benchmark (APB), a diagnostic suite of 4,209 multimodal cases across 22 domains and five evaluation settings, designed specifically to isolate planning capability from other agent skills like tool use or memory.

Testing across 12 multimodal LLMs — validated against 200 ToolSandbox and 200 τ²-bench tasks — reveals systematic weaknesses in long-horizon planning, tool-noise robustness, calibrated task refusal, and inference-time plan refinement, with APB-guided refinement improving downstream execution metrics across three representative models.

Watch APB get adopted as a standard diagnostic alongside broader agent benchmarks, and watch whether next-generation frontier models show measurable gains on the specific weaknesses the benchmark identifies.