JAZ Framework: A Minimalist AI Agent Outperforms MemGPT on Long-Horizon Recall

JAZ is an LLM agent framework built around a single core primitive called invoke, which lets the model write arbitrary executable code including recursive invoke calls, with everything visible to the model represented as variables in the code environment.

On long-horizon workflows requiring recall beyond the context window, JAZ invoke outperformed Letta (MemGPT) by 8% at half the cost on the recall-heavy portion of StuLife.

The work is a research paper, not a production-ready product, but it suggests agent systems may need less engineering scaffolding than assumed, which is worth considering for teams reviewing their current architecture.

Conventional systems rely on hand-designed tools, memory, and file systems, whereas JAZ shows that the invoke primitive alone, with prompting only, can match or exceed those systems on certain tasks.