Annus Mirabilis

Persistent-memory language models do more than recall the past. They decide which past experiences get to steer present behavior. Most memory systems rank records by relevance, recency, summaries, or tool state. But they do not ask the harder question: what will this memory make the model do next? We treat recall as an inference-time intervention. Each memory carries a causal influence signature: an estimate of how it may shift future model behavior when it enters the workspace through context, memory tools, attention, or model state. A controller admits a memory only when its expected effect, side effects, policy risk, and reliability fit the task. It then observes what happened, records lineage, and rolls back or quarantines memories whose effects diverge from prediction.

Research boundary: this is an architecture-and-analysis preprint, not a benchmark report. It defines claim boundaries, falsification tests, evaluation protocols, and deployment risks for future measurement.

Most retrieval-augmented language models choose context by query relevance. Long-lived AI systems need another signal: whether a memory is available for the current goal. We describe an activation-guided graph retrieval architecture. Request and goal cues start traversal inside an authorized memory graph. Activation and fan effects keep the search bounded. Retrieved entities emit new cues, and the system rebuilds a typed evidence graph before generation. The proposal gives five testable claims about candidate selection, traversal control, graph regularization, evidence reconstruction, and reinforcement. It also defines cost estimates, ablations, failure cases, and a reproducible evaluation protocol.

Research boundary: this is an architecture-and-analysis preprint, not a benchmark report. It defines claim boundaries, falsification tests, evaluation protocols, and deployment risks for future measurement.