KV Cache TTL for Multi-Turn Agent Scheduling

Visual companion for a podcast episode on why end-of-turn eviction works for human chat but breaks for ReAct-style agents: short tool pauses make turns tightly coupled, so evicting KV cache adds both prefix rebuild cost and queue re-entry delay.

arXiv: 2511.02230 Paper: Continuum (2025/2026 discussion) Theme: Scheduling × Memory Workload: Multi-turn agents
Main Paper
Continuum
Core Mechanism
KV Cache TTL
Scheduling Lens
Program-level FCFS
Known arXiv IDs
2511.02230
Reported Pattern
Lower job completion time
Best Fit
Pause-resume heavy agents

References

Compact source map