Action-outcome episodes
Store what the robot observed, which context was retrieved, what action was recommended or taken, and what outcome followed.
ZeptoDB
Action-Outcome Memory for Physical AI. ZeptoDB keeps robot state, sensor evidence, prior actions, outcomes, and policy context on one timeline so robotic agents can recall what worked, avoid risky repeats, and replay why a recommendation was made.
Memory for robots that act in the world
Section titled “Memory for robots that act in the world”Physical AI agents need more than chat history or a detached vector index. A robot decision depends on the physical situation: pose, sensor window, payload, zone, human proximity, policy context, the action taken, and what happened afterward.
ZeptoDB is the database layer for that memory loop. It stores high-frequency telemetry and agent memory on the same timeline, then lets an application retrieve prior action-outcome episodes with the live evidence needed to decide whether the old outcome is appropriate to reuse.
ZeptoDB is not a robot controller and does not claim certified robot safety. Its role is to provide the memory, evidence, replay, and policy-context substrate that Physical AI systems can use before and after an action.
Context-gated recall
Compare prior outcomes against current time, zone, payload, human proximity, sensor pattern, and policy context before reuse.
Replayable evidence
Join memory records back to robot state, sensor summaries, operator interventions, cache events, and tool/model calls.
Fast telemetry foundation
Ingest live robot and edge telemetry, retrieve evidence at microsecond scale, and hand data to Python without serialization.
Benchmark claims are defined by hardware, build, data shape, scope, and measurement protocol. See the benchmark criteria and reproduction notes before comparing or republishing a single number.
What the Action-Outcome research has validated
Section titled “What the Action-Outcome research has validated”ZeptoDB’s current Action-Outcome research track has moved from offline comparison into native SQL replay, edge/fleet feed replay, shadow supervisor tests, and a documented controlled-pilot boundary.
15/15 hazardous proposals suppressed
Shadow supervisor tests blocked every hazardous baseline proposal while allowing every safe context-gated proposal.
52/52 feed events acknowledged
Edge-to-fleet replay converged through bounded batches, duplicate handling, late arrivals, outage retry, and restart reload.
Native SQL replay passed
Robot state ASOF JOINs, sensor ASOF JOINs, action/outcome JOINs, suppression audit JOINs, ROW_NUMBER, LAG, and spatial checks all pass in live ZeptoDB SQL.
Controlled pilot, not GA
The runtime path is deliberately bounded: admin-gated, shadow/pilot scoped, monitored, and not described as a broad autonomous control feature.
Action-Outcome Memory → · Research evidence → · Grounded overview →
The Physical AI memory loop
Section titled “The Physical AI memory loop”Observe
Capture ROS 2 telemetry, typed robot profiles, sensor summaries, incidents, operator interventions, and edge events as ordered time-series data.
Recall
Retrieve prior episodes, reflections, policy notes, and similar action outcomes from Agent Memory with tenant, session, robot, and metadata filters.
Gate
Check whether a prior successful action still matches the current temporal, spatial, payload, policy, and human-proximity context.
Act
Let the application decide, call a model or tool on cache miss, write the decision back, and keep the full action path inspectable.
Replay
Reconstruct the chain from live evidence to retrieved memory, proposed action, suppression, outcome, and reflection.
Safety-adjacent robot workflows
Section titled “Safety-adjacent robot workflows”ZeptoDB should be described as safety-adjacent infrastructure, not as a certified safety system. It helps applications keep the evidence needed for guardrails, reviews, and cautious reuse of prior outcomes.
Risky-repeat avoidance
Suppress a prior action recommendation when the current zone, payload, sensor motif, or policy context does not match the old successful episode.
Intervention memory
Keep human/operator interventions, rollback notes, and recovery actions beside the raw robot timeline that made them necessary.
Policy evidence
Attach approval requirements, geofence context, risk tier, and allowed-action metadata to the memory record an agent retrieves.
Incident replay
Review what the robot saw, what memory was retrieved, whether a recommendation was reused or suppressed, and what happened afterward.
Available today vs research-only
Section titled “Available today vs research-only”| Status | What belongs here |
|---|---|
| Available product surface | ROS 2 scalar/profile ingest, typed profile tables, rosbag2 import/replay, time-series SQL, ASOF JOIN, Window JOIN, Agent Memory, exact/semantic cache, replay telemetry, zero-copy Python |
| Research-only evidence | Context-gated Physical AI Action-Outcome Memory experiments, synthetic robot/fleet fixtures, SQL replay harnesses, edge/fleet replay shape |
| Not claimed | Fully autonomous robot control, certified robot safety, runtime edge-to-fleet replication, or a production control plane that executes robot actions |
Read the grounded overview: Physical AI Memory and Action-Outcome Memory.
Why Physical AI teams pick ZeptoDB
Section titled “Why Physical AI teams pick ZeptoDB”- One timeline for sensor evidence and decisions. Store raw robot events and agent memory records beside the workflows that consume them.
- Memory with physical context. Retrieve by time, robot, site, zone, session, payload, metadata, TTL, importance, pinned status, recency, and access count.
- Replay instead of detached recall. A memory is useful only when the system can replay the telemetry and outcome that made it true.
- Fast enough for robot-adjacent loops. Microsecond time-series recall, millisecond-scale memory search, exact/semantic cache lookup, and zero-copy Python access.
- Open implementation path. Use GitHub, ROS 2 setup docs, Python helpers, HTTP APIs, and benchmark notes to inspect what exists today.
Built for Physical AI and adjacent systems
Section titled “Built for Physical AI and adjacent systems”Physical AI & Robotics
ROS 2 typed profiles, rosbag2 replay, sensor fusion, and action history become replayable episodes. A robot agent can recall what it saw, what it tried, and which previous intervention worked.
Industrial Robots & Smart Factory
Factory telemetry becomes operational memory. Maintenance and quality agents can combine sensor history, work orders, operator notes, and safety procedures.
Autonomous Systems
Fleet logs, scenario mining, disengagement notes, and agent decisions stay on one timeline for scenario replay and evidence review.
Logistics & Edge Automation
AGV pose, RFID, sorter, cold-chain, and yard events become spatially queryable timelines for replay, geofencing, and exception agents.
AgentOps for Embodied Agents
Store retrievals, cache events, model calls/errors, context traces, replay windows, tool calls, and outcomes as ordinary time-series tables.
How it compares
Section titled “How it compares”| Vector DB only | Generic TSDB | Prompt cache only | ZeptoDB | |
|---|---|---|---|---|
| Robot telemetry timeline | separate | native | no | native time-series |
| Action-outcome memory | text-like recall | manual schema | no | memory beside evidence |
| Temporal joins | no | varies | no | ASOF + Window JOIN |
| Policy/context gates | application-only | application-only | no | queryable evidence + memory metadata |
| Replay path | detached from sensors | telemetry only | response only | sensor → memory → action → outcome |
| Python path | serialized | usually serialized | n/a | 522ns zero-copy |
Detailed comparisons: Agent Memory vs Vector DBs · vs kdb+ · vs ClickHouse · vs InfluxDB · vs TimescaleDB
Start building
Section titled “Start building” Physical AI & Robotics ROS 2 ingest, typed profiles, rosbag2 replay, action memory, and safety-adjacent replay
Action-Outcome Memory Physical AI action, recovery, suppression, and outcome memory
Action-Outcome Evidence Experiment 013-023 results, runtime maturity, and controlled-pilot boundaries
Physical AI Research Overview What ZeptoDB has actually validated, with research-only limits kept clear
ROS 2 Setup Build and smoke test the ROS 2 connector path
Agent Memory Memory records, context retrieval, cache lookup, and replay for agents
Python Quickstart Copy-paste memory, context, cache, and write-back flow
Benchmarks Criteria, reproduction notes, ingestion, query latency, Python zero-copy, and memory recall
GitHub Read the source, file issues, and follow community discussions