.aamt – one GPU file format for model weights, agent memory and worlds (asamanthinks.com)

The introduction of the .aamt file format marks a significant advancement in GPU-based AI applications, paralleling what GGUF offered for CPUs. Designed to encapsulate not only model weights but also agent memory and environmental context, .aamt provides a unified structure for managing complex data in real-time scenarios. This format allows for the seamless incorporation of dynamic thoughts and observations into a spatial representation, enabling models to compute environments and interactions live. A standout feature is the use of geometry for memory recall, which eliminates the need for traditional embedding servers, optimizing efficiency by addressing memories directly through a spatial coordinate system. For the AI/ML community, .aamt is particularly notable for its deterministic packing and built-in error correction mechanisms, ensuring data integrity and consistency across different machines. With features such as a hybrid logical clock for synchronizing edits and two copies of each memory for redundancy, .aamt enhances collaboration and data reliability in distributed environments. Moreover, its versatile support across multiple programming languages and environments, including native C++, Python, and JavaScript, positions it as a crucial tool for developers working with complex AI systems that require robust memory and data handling capabilities.