跳到主内容
精选85elvis论文研究

阿里新方法:把Agent上下文管理当编程任务

Impressive work from Alibaba.

原文
推荐理由

做长时运行Agent的同学必看,这套把上下文管理当编程任务的做法在长记忆基准上大幅领先,值得照着事件日志+持久内核的思路改造自己的记忆层。

Impressive work from Alibaba.

阿里巴巴的出色工作。

(bookmark it)

(收藏它)

If you build long-running agents and keep rewriting your memory schema, take a look at this approach.

如果你构建长期运行的代理并不断重写记忆模式,不妨看看这种方法。

It basically treats agent context management as a programming task.

它基本上将代理上下文管理视为一项编程任务。

Here is how it works:

其工作原理如下:

It backs each agent session with an append-only event log and a sandboxed, persistent Python kernel.

它为每个代理会话配备了一个仅追加的事件日志和一个沙盒化的持久Python内核。

Tool outputs, retrieved history, and derived state bind to typed variables across model calls instead of being serialized into the prompt every turn.

工具输出、检索到的历史和派生状态在模型调用之间绑定到类型化变量,而不是每轮都序列化到提示中。

Model-written code searches and transforms that state, and only explicitly printed projections enter the working view.

模型编写的代码搜索并转换该状态,只有明确打印的投影才进入工作视图。

The event log keeps lossless ground truth, so nothing has to be committed to a compressed form before you know what will matter later.

事件日志保留了无损的原始数据,因此在你知道哪些信息以后会重要之前,无需将任何内容压缩成简化形式。

When the working view nears its budget, stale spans are evicted but stay recoverable. An eviction index keeps compact landmarks tied to exact event-log addresses, so the agent navigates straight back to a region instead of searching the whole log.

当工作视图接近其预算时,过时的片段会被淘汰但仍可恢复。淘汰索引保留与精确事件日志地址关联的紧凑地标,使代理能直接导航回某个区域,而不是搜索整个日志。

Results: with Qwen3.8-Max, 94.8% on LongMemEval_S, 73.1% on BEAM_10M (5.1 points over the best published memory system), and 86.7% on LOCA_256K.

结果:使用Qwen3.8-Max,在LongMemEval_S上达到94.8%,在BEAM_10M上达到73.1%(比最佳已发表记忆系统高出5.1个百分点),在LOCA_256K上达到86.7%。

Treating context management as a programming task means it inherits every future improvement in model coding ability.

将上下文管理视为编程任务意味着它继承了模型编码能力未来的一切改进。

Paper: https://arxiv.org/abs/2608.21690

论文:https://arxiv.org/abs/2608.21690

Track more trending AI papers in our academy: https://academy.dair.ai/

在我们的学院中追踪更多热门AI论文:https://academy.dair.ai/

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

进入量化体系 →

相似阅读

另一事件,读法相近