MIT open-sources JAZ framework built on a single invoke primitive
A minimal prompt-and-loop design outperforming heavyweight memory infrastructure suggests agent scaffolding is largely redundant overhead, which changes how teams should budget for agent architectures. Developers building long-horizon agents can study the JAZ repo and test whether dynamic scoping replaces their vector-store and meta-optimisation layers. Benchmarks come with explicit dollar costs, making direct comparison with Letta (MemGPT) and ACE easy to replicate.
MIT released JAZ, an open-source agent framework that replaces layered memory and optimisation add-ons with one invoke function plus a plain prompt loop.
On the StuLife long-range recall benchmark it scored 69.9% at a cost of US$18.
3, beating Letta (MemGPT), a dedicated memory framework, which scored 61.8% at US$42.
1.On the AppWorld self-evolution test it reached 74.
2% versus 69.9% for ACE.
The design treats prompts, tools and interaction history as ordinary Python variables that models read and write, and spawns sub-agents that inherit full session state when context runs low.
MIT开源JAZ框架:用极简invoke(调用原语)终结复杂Agent外挂
现在做长程记忆和自我进化的Agent,行业习惯套上一层又一层的复杂外挂系统,架构越来越臃肿。MIT团队提出的JAZ打破了这个范式,仅仅依靠纯提示词和一个极简循环,在长程基准StuLife远距召回评测中以$18.3成本跑出69.9%通过率,超越成本$42.1的专用记忆框架Letta(MemGPT)的61.8%;在AppWorld自我进化测试里也以74.2%的成绩压制了经典外挂系统ACE的69.9%。
在JAZ中,你不再需要专门搭建向量记忆库或复杂的元优化管道。你只需使用一个统一的invoke函数,让大模型在Python环境中把提示词、工具和自身交互历史全部当成普通变量读写;当上下文即将耗尽时,你可以让它直接生成子Agent,把完整的会话变量引用无损移交下去,或者让顶层Agent把自身当成优化器,动态批处理调优下层任务。
这套机制的底座是把Agent执行本身视作动态编写函数体的语言原语,并利用动态作用域(Dynamic Scoping)让工具与监控钩子(Hook)自动向下渗透到所有递归子节点。这意味着历史记录不再是只能看不能碰的字符串,而是能被代码直接切片、搜索和传递的变量,长记忆与自迭代自然内生于控制流中。
Github:https://t.co/4qFuvzWuKK
This report is categorized under AITOP's "Models" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows and US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include JAZ, Agent, MIT, open-source models.
Source attribution: curated from Gorden Sun (@Gorden_Sun).