Retrieval‑augmented language model systems usually select context based on query‑conditioned relevance. We argue that long‑lived AI systems need a second operator: activation‑conditioned memory availability. This paper specifies an architecture where:
- Request and goal cues seed an authorized memory graph.
- Activation and fan attenuation allocate a bounded traversal frontier, keeping the search tractable.
- Retrieved entities recursively emit cues, driving further graph expansion.
- A typed evidence graph is reconstructed before generation for the model to consume.
The specification yields five architectural consequences: candidate selection, traversal control, graph regularization, evidence reconstruction, and reinforcement interaction. We provide implementation‑level cost estimates, define ablations, failure conditions, and a reproducible evaluation protocol. The central claim is narrowed or falsified if a simpler graph reranker matches full traversal on causal and temporal recovery under equal budgets. No benchmark measurements are reported. If the claim holds, retrieval returns the evidence path that makes a memory available, not merely a ranked list.
Review: The work offers a systematic graph‑based view of memory activation, laying out clear evaluation criteria despite the current lack of empirical results, and thus paves the way for future implementations.