Remaneta Treats an AI Conversation as Hardware

Field notes

The proposed memory controller tracks session identity, lifecycle state, idle time, turn count, and placement across three tiers, allowing an agent conversation to be demoted without a host command. Remaneta describes restoring a 4.1 MB per-token index before reading from a much larger 440 MB session, and reports 0.22 ms p99 simulated wake latency. The page also lists RM-1 serving, RM-T training, and RM-E edge configurations, while noting that its package diagrams are schematic.