Dossier 05 · Robotics, the frontier
A memory for the machines that will walk among us.
Robotics is young. That is exactly when its memory should be designed. A robot cannot outsource its memory to a cloud. Latency breaks reflexes, a lost connection cannot mean amnesia, and what a machine remembers about your factory, your home or your streets has to stay on hardware you control, auditable and answerable to a human.
Mnemosyne OS is that memory layer by construction. Encrypted vaults on the machine itself, semantic search with no network, and a human who decides what is kept and what is forgotten.
Robotic memory : on-board memory for robots and autonomous systems: local, encrypted, network-independent, and governed by the human. Mnemosyne OS is a memory engine built to live where the robot lives.
None of this is a robotics product today. It is where we are heading. We are looking for robotics teams, industrial or academic, to explore it with us.
Our thinking How we picture it: memory paced by the motors
A design thesis, not a shipped feature. The pieces below already run in Mnemosyne OS on a desktop. What follows is how the same architecture would sit on a machine that moves, paced by the state of its motors.
Motors running. In active operation the memory engine stays out of the control loop. No deep consolidation, no re-indexing: the compute belongs to perception, navigation and reflexes. Memory only answers, and it answers fast, from the vault nearest the task. The whole database never wakes.
Motors at rest, charging. This is when the dream engine wakes, the same consolidation Mnemosyne OS already runs on downtime. It ingests the day’s raw streams, drops the noise, recomputes embeddings and lets new links surface. The machine’s version of hippocampal replay: the step from recording to understanding, done when nothing else needs the compute.
Spines, not one big brain. Memory is split into isolated, encrypted vaults on the machine: spatial navigation apart from user profiles, by domain and by criticality. Retrieval happens in the relevant vault, without waking the whole store. Those vaults and this spine structure exist in the OS today. Giving each one its place on a robot is what remains.
To be clear: Mnemosyne OS runs on desktops today. No robot deployment, no on-device NPU integration, no benchmark on hardware that moves. This is the map we would follow with a robotics partner.