Telemetry and memory together
Keep robot state, sensor summaries, decisions, tool calls, interventions, and outcomes on a shared temporal substrate.
ZeptoDB
A time-series and memory database for systems that act in the physical world. Keep robot state, prior actions, outcomes, and the evidence needed to decide whether an old result still applies.
Memory for systems that act
Section titled “Memory for systems that act”A robot does not need another detached history store. It needs a record that keeps the physical situation, retrieved context, proposed action, observed outcome, and later reflection attached to the same event-time evidence.
ZeptoDB provides that substrate. High-frequency telemetry and agent memory remain queryable on one timeline, so an application can inspect what happened before reusing a previous outcome. ZeptoDB supplies memory and replay; the application, controller, model, or operator remains responsible for policy and execution.
Context before reuse
Check time, zone, payload, sensor pattern, policy, and human proximity before treating an old success as relevant now.
Replayable decisions
Reconstruct the path from observed state to retrieved evidence, proposed action, suppression, and eventual outcome.
Performance claims depend on hardware, build, data shape, and measurement protocol. See the benchmark criteria and reproduction notes for the measured scope.
Observe, recall, gate, act, replay
Section titled “Observe, recall, gate, act, replay”- Observe. Ingest ordered telemetry, typed robot profiles, incidents, and operator interventions.
- Recall. Retrieve prior episodes and outcomes with tenant, session, robot, time, and metadata filters.
- Gate. Compare the old result with the current temporal, spatial, operational, and policy context.
- Act. Let the surrounding application decide, then write the recommendation or intervention back to the timeline.
- Replay. Inspect the complete evidence chain after the outcome is known.
The Action-Outcome Memory overview describes the episode model. The Physical AI workflow shows where ROS 2 ingest, sensor replay, and operator review fit.
Results, not just a roadmap
Section titled “Results, not just a roadmap”The experiments deliberately move from an offline comparison to native SQL replay and then to bounded runtime fault cases. These are the three milestones that best explain the direction.
12/12 proposals converged after injected projection faults
Can fresh runtime objects repair committed state without duplicating decision or evidence rows?
- Proposals converged
- 12/12
- Faults repaired
- 3/3
- Duplicate ids
- 0
The commit ledger held as an effectively-once boundary for the bounded supervisor sink.
Boundary:Supervisor-specific contract; not a generic multi-table SQL transaction.
15/15 hazardous proposals were suppressed in shadow replay
Can a shadow supervisor suppress hazardous baseline proposals and remain idempotent after restart?
- Hazardous suppressed
- 15/15
- Safe allowed
- 5/5
- Restart skips
- 20/20
The experiment connected historical outcomes to explicit allow, suppress, and manual-review decisions.
Boundary:Research-only shadow fixture; no autonomous actuation.
Context gating avoided every risky repeat in the fixture
Can a robot reuse a prior recovery without repeating an action that only worked under a different physical context?
- Recovery Top-1
- 1.00
- Risky-repeat avoidance
- 1.00
- Hazardous Top-1
- 0.00
The result moved the problem from incident search to evidence-backed action reuse.
Boundary:Five synthetic Physical AI incident families; not a production safety guarantee.
See the complete experiment timeline →
Available now and still experimental
Section titled “Available now and still experimental”Available product surface
Time-series SQL, ASOF and Window JOINs, ROS 2 scalar and typed-profile ingest, rosbag2 import and replay, Agent Memory, prompt cache, replay telemetry, and zero-copy Python paths.
Experiment and pilot surface
Context-gated Action-Outcome fixtures, edge-to-fleet feed semantics, and the shadow supervisor runtime remain research or bounded experimental paths. They are not certified robot safety or autonomous control features.
Choose the next layer
Section titled “Choose the next layer” Understand the product Current database capabilities, storage, ingestion, clients, security, and deployment
Inspect the experiments Result-first summaries, limitations, methods, and raw reproducible evidence
Start building Install ZeptoDB, run the quick start, and follow the API and ROS 2 guides