Remaneta Contact
01 · Agent serving

Long-lived agents at fleet scale

A tool-using agent alternates short bursts of computation with long waits. Measured production traffic shows a minority of calls consuming most of the wall-clock, with idle intervals spanning three orders of magnitude.

Under SCMA a waiting agent descends to cheaper memory and returns inside the wake objective, so a fleet is sized by the agents it holds rather than the few it is actively talking to.

02 · Reinforcement learning

Rollout trees that branch

Rollouts dominate reinforcement-learning wall-clock, and rollout trees branch repeatedly from a common prefix. Copying that prefix for every branch is pure waste.

Hardware copy-on-write fork lets a branch reference its parent's state rather than duplicate it, including while parked in reduced-precision form. Bit-exact determinism makes a run reproducible.

03 · Confidential inference

Encrypted at rest by construction

Session state is among the most sensitive data a serving system touches; it is the user's conversation. Encrypting it usually means an extra pass someone has to schedule.

Here encryption happens as a side effect of leaving the active tier, under a key belonging to that session, in the same traversal that moves and requantises it. The hash-chained journal then gives an auditable record of where state has been.

04 · Edge and on-device

Few sessions, held a very long time

An assistant on a device holds a small number of sessions across days, through sleep and power events, on a strict energy budget.

The same lifecycle applies with a reduced state set: a session persists in non-volatile memory in compressed form and resumes without recomputation. This is the intended scope of RM-E and the licensable Profile-E core.