Planner context in, outcomes out.
A robot bridge labels each episode by embodiment, site and skill, asks the gateway for context, and reports the outcome. If the network drops, a signed durable outcome queue survives restarts and retries with the same occasion, and conflicting final outcomes are refused.
Because outcomes flow back, the same randomized evidence that measures an agent’s memories measures a robot’s: which recovery hints, site notes and skill lessons actually improve success for this embodiment at this site.
from unimem.robotics.bridge import HTTPGateway, RobotBridge
from unimem.robotics.edge import OutcomeQueue
robot = RobotBridge(
HTTPGateway("http://127.0.0.1:8787", token),
embodiment="parallel-jaw", site="depot", skill="grasp", robot="robot7",
outcome_queue=OutcomeQueue("/var/lib/unimem-edge"),
)
context = robot.start("episode-42", "Recover the dropped object")
robot.finish("episode-42", succeeded=True)
Working memory inside the control loop.
In-process working memory holds bounded vectors and keyframe metadata with local cosine ranking and a 5 ms default budget. It returns the best result found so far, marks overruns, and makes no network or persistence calls in the retrieval path.
Safety keyframes stay protected under eviction. At episode end only predictive keyframes are consolidated into long-term memory, with lineage and an append-only audit record.
5 ms is a design target with a separate wall-clock benchmark gate. Integrators must benchmark their own hardware and vector dimensions.
A shield between memory and motion.
Every proposed memory-derived action can be checked against a bounded Signal Temporal Logic specification registered to the memory that suggested it. Missing memory, missing or out-of-order telemetry, and negative robustness hold the action. The bridge fails closed when no shield or no recalled memory is present.
The shield never sends motor commands. It complements, and never replaces, the robot’s safety-rated controller, emergency stop and risk assessment.
Memory-to-action shieldWhat one robot learns, the site can know.
Collective site memory ingests a VDA 5050 subset and an Open-RMF adapter boundary into vendor-neutral, bitemporal site events. Scene graphs keep sensor source, event time and record time; conflicting observations stay visible instead of being silently merged.
A signed mesh lets robots share memory peer to peer. Conflicting concurrent values become explicit contradictions, safety additions cannot be removed, and forged or revoked peers are rejected. Signed edge updates carry protected hazards and proven facts with Merkle consistency proofs.
Collective site memoryLiving safety cases from real evidence.
A digest-bound claim and evidence graph tracks which safety claims rest on which memories and outcomes, flags evidence that has gone stale, and mines reproducible scenarios, so a safety case can change as the fleet learns rather than drift silently out of date.
These are engineering aids with stated limits. They are not certification, and robot motion and routing stay under the fleet’s certified control stack.
Living safety casesBring one robot workflow.
Try the local loop, or bring one workflow to a pilot conversation.
Technical details link to the inspected source. Check deployment requirements before enabling optional features.