Memory Harnesses for Long-Running Research Agents — Stefania Druga, Sakana.ai
Memory harnesses for local models can solve context rot in long-horizon research agents.
“that makes this issue of dealing with context rot a priority”
Sakana AI researcher Stefania Druga presents a practical framework for managing context degradation in long-running agents using local models like Qwen 27B and DeepSeek V4 Flash on consumer hardware. The talk frames context rot as an urgent problem as agent tasks grow longer while model release cadence slows, creating a convergence point. The approach emphasizes local model inference with structured memory harnesses as a cost-effective alternative to cloud APIs, citing Coinbase's success reducing AI spend through better caching and routing.