Long-horizon robotic tasks expose a fundamental weakness in current Vision-Language-Action (VLA) models: their restricted temporal context prevents them from effectively accumulating and reusing past experiences. A new survey paper addresses this by systematically reviewing memory mechanisms that could enable embodied agents to reason over extended periods and adapt their decisions.

The authors propose a taxonomy based on memory scope and persistence, grouping existing approaches into five representative paradigms and one emerging frontier. They then compare representative VLA memory systems across memory characteristics, architectural complexity, and open-source maturity. A clear trade-off emerges: as architectural complexity increases, open-source maturity generally decreases.

The survey also highlights key challenges for building full-lifecycle memory systems, including efficient memory management and lifelong knowledge evolution. By offering a unified perspective on memory architectures, the authors aim to guide future development of lifelong embodied intelligence.