Research Unpacks Agent Memory: The Memory Contract of Sparse Event-KV
Zefeng Cai · hf · 2026-08-04
This paper explores how long-horizon agents reuse their KV cache as memory, introducing the concept of the memory contract for sparse event-KV serving.
The authors test whether a retained event remains informative once the observations that produced it are dropped. They identify semantic materialization: a downstream event's cached rows act as an independently servable view of computation whose inputs are gone. A deliberately phrased, answer-free event can raise donor-aligned recovery from 6% to 51% on Qwen3-8B without ever naming the value. The findings reveal that compact state survives while larger payloads decay toward chance. Consequently, dropping a source event and observing no accuracy loss does not prove the source was unnecessary.
More from coding & agent
- Karpathy Tests Claude Opus: Writes 5,500 Lines of Code in 2 Hours to Generate 3D Film — iamrobotbear · 2026-08-04
- Automating Agent Optimization: Devs Share Recursive Auto-Improvement Workflows — iamrobotbear · 2026-08-04
- Easy to Prototype, Hard to Maintain: OpenHands Shares Coding Agent Pitfalls — rajistics · 2026-08-04
- OpenBMB Launches Dual-Agent System to Automate Supercomputing Acceleration — aigclink · 2026-08-04
- Cloudflare Plans to Launch a Wallet Tailored for AI Agents — vista8 · 2026-08-04
- 76 Production-Ready Skills Released to Expand Claude's Capabilities — tom_doerr · 2026-08-04