Heshware is investigating whether a dynamically activated graph of memory nodes can serve as an interpretable long-term memory layer for a local language model. This proposal stems from the Marven Neural Node Network, originally an interactive visualization of cognitive processing.
The proposed architecture includes ingestion of structured memory objects, vector retrieval for semantic similarity, and a graph structure for relationships and temporal links. It also features activation values for context entry, decay processes for consolidation, and a sandboxed synthesis process to recombine memories without treating them as facts.
The author seeks technical feedback on node activation metrics, the suitability of graph neural networks, and methods to combine vector similarity with explicit graph relationships. The goal is to develop a testable architecture that improves retrieval accuracy, continuity, and interpretability while resisting memory poisoning.