Tencent has released the ContextPilot-8B checkpoint, based on Qwen3-8B, as part of a framework designed to teach agents proactive context management for long-horizon tasks. The system enables agents to plan, maintain long-term memory, and offload less useful context while reasoning and using tools.
ContextPilot combines three main components: an extended toolset with planning, structured memory, retrieval, and soft context offloading; context-aware partial rollout focusing on sensitive context-editing decisions; and fine-grained credit assignment training intermediate snapshots using downstream outcomes. The framework addresses limitations in existing methods by supporting global planning and adaptive compression while improving exploration efficiency.
Experiments on long-context question answering and deep search tasks show that ContextPilot achieves stronger performance with a more compact working context, consistently outperforming existing baselines across various base models and benchmarks.