Papers

Persistent-memory language model systems increasingly decide which prior user, project, or organizational records should influence present inference. Existing memory and retrieval systems usually expose relevance, recency, summarization, or tool-mediated persistence; they do not by themselves specify how a recalled memory's downstream behavioral effect should be predicted, constrained, verified, and reversed. We formulate memory recall as a controlled inference-time intervention. Each memory entity is associated with a causal influence signature: an estimate, in a canonical influence space, of the directional change in future model behavior produced by applying that memory through context, an external memory interface, attention, or latent state. A controller selects admissible memory interventions by target effect, side-effect budget, policy risk, and reliability; observes the realized effect; records lineage; and rolls back or quarantines interventions whose observed effect diverges from prediction. The contribution is an architecture-and-analysis proposal, not an empirical benchmark report. We state claim boundaries, falsification conditions, evaluation protocols, and deployment risks for future measurement.

Retrieval-augmented language model systems generally select context by query-conditioned relevance. We argue that long-lived AI systems require a second operator: activation-conditioned memory availability. We specify an architecture in which request and goal cues seed an authorized memory graph, activation and fan attenuation allocate a bounded traversal frontier, retrieved entities recursively emit cues, and a typed evidence graph is reconstructed before generation. The specification yields five testable findings about candidate selection, traversal control, graph regularization, evidence reconstruction, and reinforcement. We derive implementation-level cost estimates and define ablations, failure conditions, and a reproducible evaluation protocol. This architecture-and-analysis preprint reports no benchmark measurements.