MIT's JAZ agent framework beats Letta on long-range recall at under half the cost

JAZ, open-sourced at jaz-lang/jaz, gives the model a single invoke primitive in a Python REPL where its prompt and its own history are variables, so it can pass its full history by reference to a subagent when the context window fills up. With GPT-5.4 nano it passed 70% of StuLife's far-recall tasks vs 62% for Letta (MemGPT) at about $18 vs $42 per run, and on AppWorld self-improvement it hit 74% vs ACE's 70% at lower cost.