Robots are getting smarter, but they still suffer from a terrible case of amnesia. While modern vision-language models can see and react in the moment, they struggle to remember what happened yesterday or plan for next week. This lack of long-term memory is the bottleneck preventing robots from handling complex, real-world tasks that require patience and context.

Researchers have introduced ABot-AgentOS, a new operating system layer designed to give robots persistent memory and reasoning capabilities. Sitting above basic movement controls, this system acts as a deliberative brain. It manages planning, verifies actions, and uses a unique Universal Multi-modal Graph Memory. This memory system converts everything the robot sees, hears, and does into a structured graph of nodes and edges. It tracks dialogue, visual observations, spatial context, and time, creating a detailed history that the robot can reference to improve future performance.

The system includes a self-evolution loop that learns from its mistakes. When the robot fails, the system diagnoses the error and creates a learning asset. Crucially, this improvement is gated, meaning it only applies to future tasks. This prevents the robot from cheating by using answers from the current test set, ensuring that the learning is genuine and the evaluation is honest. To test this, the team created EmbodiedWorldBench, a benchmark with over 200 tasks across indoor, outdoor, and hybrid scenes, covering everything from object search to dynamic events.

The results show that this general agent layer significantly boosts performance. ABot-AgentOS outperforms single-controller baselines in both task success and goal completion. It also sets new standards on memory benchmarks, achieving high scores on tests like LoCoMo and Mem-Gallery. The self-evolution feature further pushes these numbers up, proving that robots can genuinely improve over time. This research suggests that giving robots a structured, auditable memory is the key to unlocking long-horizon autonomy.