Could a Graph-Based "Neural Node Network" Become an Interpretable Memory Layer for a Local AI?

A calibrated one-hop graph reranker produced identical hit-at-five, MRR, and trusted evidence sets to graph-free retrieval across 250 frozen paired requests, with graph candidates appearing in only 5 requests and none becoming new trusted evidence, according to the author's RE-call benchmark. The author reports the raw transcript benchmark contained zero authored graph relations versus 44,364 in the real RE-call memory corpus, arguing a node set without typed relationships cannot meaningfully test graph reasoning. The evaluation design is documented in the Evidence Graph V1 preregistration on GitHub, and the author recommends typed edges, bounded propagation, and shuffled edge controls while postponing a GNN until deterministic graph retrieval shows a repeatable structural failure with a useful training signal.

This is very close to a direction I have been testing . My main conclusion is that the graph should be an interpretable projection over canonical evidence, not the sole source of truth. Memories should retain their source, validity interval, lineage, and trust status. Graph expansion can propose related evidence, but those results still need to pass provenance, currentness, citation, and abstention checks. Activation should also remain separate from truth. It indicates associative relevance, not confidence, authority, disclosure permission, or permission to act. Observed, derived, inferred, synthesized, and external memories should remain distinguishable. I recently tested a calibrated one hop graph reranker. In a frozen paired run, graph retrieval and graph free retrieval produced identical hit at five, MRR, and trusted evidence sets across 250 requests. Graph candidates appeared in only five requests, and none became new trusted evidence. This does not disprove graph memory. It shows that the query set did not contain enough graph solvable cases. I also found that my raw transcript benchmark contained zero authored graph relations, while the real RE-call memory corpus contained 44,364. A node set without typed relationships is not a meaningful test of graph reasoning. I would therefore begin with typed edges, bounded propagation, and shuffled edge controls. I would postpone a GNN until deterministic graph retrieval demonstrates a repeatable structural failure with a useful training signal. The evaluation design is documented in the Evidence Graph V1 preregistration https://github.com/GiulioDER/RE-call/blob/master/benchmarks/PREREGISTRATION-evidence-graph-v1.md .