Rethinking Memory as Continuously Evolving Connectivity

Researchers introduced FluxMem, a memory framework that evolves a heterogeneous graph to help AI agents store and retrieve memory more adaptively.

The system updates memory through formation, feedback-driven refinement, and long-term consolidation, aiming to better handle dynamic multi-step tasks.

FluxMem reports state-of-the-art results on LoCoMo, Mind2Web, and GAIA, outperforming prior approaches such as LightMem.

The authors say the code will be open sourced.