JAZ outperforms Letta and ACE in long-term memory tasks

ACE: ACE stands for Agentic Context Engineering, a framework that enables LLMs to self-improve by treating context as evolving playbooks updated through generator, reflector, and curator roles. It focuses on incremental, structured updates to instructions rather than model fine-tuning. The news highlights JAZ achieving better results than ACE on continual self-improvement tasks using only a minimal loop. JAZ: JAZ is a minimalist LLM agent framework developed by researchers at MIT CSAIL and introduced in a September 2026 arXiv paper. It centers on a single 'invoke' primitive that lets the model write executable code, recursively call subagents, and treat its own prompt and full interaction history as accessible code variables. In the reported work, JAZ demonstrates that a basic agent loop can handle long-term recall and self-improvement tasks without dedicated external memory or tuning systems. Letta: Letta is an open-source platform for building persistent AI agents that maintain continuous memory, identity, and skills across sessions and model changes. It evolved from MemGPT research and emphasizes structured context management and dynamic workflows. The news positions JAZ as a simpler alternative that outperforms Letta on recall-intensive benchmarks. Framework Design: JAZ generalizes code-mode agent loops by making the LLM's interaction history and inputs directly manipulable as code variables. Open Source Availability: The JAZ framework and evaluation code have been released on GitHub following the paper's publication.