Remembering Through Change: Provenance-Preserving Hypergraph Memory for Multimodal Agents
MHG-Mem reports an 8.20–9.98% performance improvement over the strongest baseline across three benchmarks.
The proposed memory system links entity identities and changing states to their original visual evidence. It groups observations of the same entity, then builds aspect-specific accounts of how that entity evolves across sessions. Its search method follows typed links through the memory and checks source observations when fine visual detail needs verification. This supports both visual checks and reasoning about changes between sessions.
ArXiv · AI/CL/LG's note
The proposed memory system links entity identities and changing states to their original visual evidence. It groups observations of the same entity, then builds aspect-specific accounts of how that entity evolves across sessions. Its search method follows typed links through the memory and checks source observations when fine visual detail needs verification. This supports both visual checks and reasoning about changes between sessions.
ArXiv · AI/CL/LG's note
score 4