Tag
This paper introduces agent-controlled forgetting for tool-using agents, where the model replaces bulky tool observations with short recoverable notes, cutting prompt tokens by ~50% and API costs from ~$4.38 to ~$1.28–$1.44 in an exploratory debugging case, while highlighting workload-dependent tradeoffs such as more requests and longer runtime.
MemPilot is a framework that uses reinforcement learning to dynamically curate multimodal memory for LLM agents on demand, balancing performance, cost, and latency via adaptive retrieval and heterogeneous model selection. Experiments across five benchmarks show it achieves better performance-cost-latency trade-offs than existing trade-off-aware baselines.
This paper introduces Just-in-Time Memory (JitMem), a method for LLM agents that defers memory curation to read-time for task-adaptive payloads, demonstrating significant performance improvements over baseline methods in benchmarks like ALFWorld and WebShop.