New agent framework tops specialized memory systems, paper says
JAZ, built around a single 'invoke' primitive with no dedicated memory module, beat Letta and ACE on two agent benchmarks, according to a paper posted to arXiv.
A team of researchers built a minimalist AI agent framework called JAZ around one primitive, a new arXiv paper says. It says the framework beat two specialized systems on benchmark tasks.
https://x.com/omarsar0/status/2103826930181308720
The framework exposes one operation, "invoke," letting a model write executable code and call itself recursively. Every prior input and result stays available as a variable in that code, the paper says.
On StuLife, a recall-heavy benchmark, the paper reports JAZ beat Letta, the MemGPT-based memory system, by 8% at half the computing cost. On AppWorld, a benchmark for continual self-improvement, it says JAZ beat the ACE method by 4% while costing less to run.
The AI newsletter writer omarsar0 flagged the paper, saying it came from MIT's Computer Science and Artificial Intelligence Laboratory. He urged developers to "keep your agent harness minimal." The authors reported results on only these two benchmarks, and no independent group appears to have verified them yet.