AMV-L: Lifecycle-Managed Agent Memory for Tail-Latency Control in Long-Running LLM Systems

This collection of arXiv papers (Mar 6 2026) covers systems and ML-serving research with several operational and security-relevant findings. Key contributions: AMV-L proposes value-driven lifecycle-managed agent memory for long-running LLM agents, bounding retrieval candidate sets to avoid heavy-tailed latency and dramatically improving throughput and tail latency versus TTL and LRU retention schemes. Papers on Prefill-Decode disaggregation and PromptTuner present SLO-aware models for allocating prefill/decode resources and elastic prompt-tuning services, tying resource counts to latency SLOs,