阿里 Scroll：让模型写程序管理上下文，长程 Agent 不再丢历史
What if context management were a program the model writes rather than a rule yo…
What if context management were a program the model writes rather than a rule your harness applies?
如果上下文管理是模型编写的程序，而非你的工具链应用的规则，那会怎样？
Compressing an agent's history to fit the context window is what breaks long-horizon agents, so keep the raw log intact and let the model write code to pull back exactly what it needs when a question comes up.
压缩代理的历史记录以适应上下文窗口，正是破坏长周期代理的因素，因此保留原始日志完整，让模型编写代码，在问题出现时精确提取所需内容。
The problem is timing. Summarization and memory extraction both decide what to keep at write time, before anyone knows what the agent will be asked later. Whatever the summary drops is unreachable, even though the full log is still sitting on disk.
问题在于时机。摘要和记忆提取都在写入时决定保留什么，那时还不知道代理之后会被问什么。摘要丢弃的内容无法访问，即使完整日志仍存储在磁盘上。
Scroll, from Alibaba, moves that decision to query time. The session lives outside the prompt in an append-only event log where every turn has a stable address, plus a sandboxed Python kernel that persists across model calls. Tool outputs get bound to variables in the kernel instead of pasted into the prompt.
阿里巴巴的Scroll将这一决策移至查询时。会话存在于提示词之外，采用仅追加的事件日志，每个轮次都有稳定地址，外加一个跨模型调用持久化的沙箱Python内核。工具输出绑定到内核中的变量，而非粘贴到提示词中。
With Qwen3.8-Max as the backbone, it reports 73.1 on BEAM histories current context windows cannot hold, against 68.0 for the best published memory system on its own setup.
以Qwen3.8-Max为骨干，它在当前上下文窗口无法容纳的BEAM历史记录上报告73.1分，而自有设置下最佳已发表记忆系统为68.0分。
– arxiv. org/abs/2608.21690
– arxiv.org/abs/2608.21690
Title: "Context as an Environment: Programmatic Context Management for Long-Horizon Agents"
标题：“上下文作为环境：长周期代理的程序化上下文管理”
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力