MIT’s JAZ Framework Outperforms Letta and ACE in AI Memory Tasks

Researchers at MIT CSAIL have unveiled JAZ, a minimalist agent framework that treats an AI's history as code variables. By allowing models to write executable code to manipulate their own past, JAZ aims to replace complex, tiered memory systems with a more efficient approach.

In testing on the StuLife benchmark using the GPT-5.4 nano model, JAZ achieved a 70% score in long-term recall, outperforming Letta, which scored 62%. Notably, JAZ reached these results at approximately half the cost of Letta's stateful memory hierarchy.

The framework also demonstrated superior performance in self-improvement tasks. Using the AppWorld benchmark, JAZ scored 74%, surpassing the ACE harness by 4 percentage points while maintaining lower operational costs.

The research, titled Harness as a Language: A Minimalist Agent Framework With Maximal Expressivity, is available on arXiv under identifier 2609.26891. The team has released the framework and evaluation code on GitHub under the repositories jaz-lang/jaz and jaz-lang/jaz-evals.