Action-Outcome Memory

Physical AI agents need more than perception and fast telemetry. They need memory of what was tried, what happened afterward, and whether the same action should be repeated, suppressed, or escalated in the next similar context.

ZeptoDB is building that as Action-Outcome Memory: a time-series memory plane for robot operations, edge safety decisions, fleet audit, and policy improvement.

Every action becomes evidence. Every outcome becomes memory. Every future decision can be checked against what actually happened before.

The problem

Robots, drones, mobile fleets, cold-chain systems, and industrial agents increasingly make operational decisions in real environments. A generic log can show that an action happened. It usually cannot answer the higher-value question:

Should this action be repeated in this context?

That requires keeping the full sequence intact:

robot state before the decision,
sensor summaries near the decision,
retrieved historical incidents,
selected or suppressed actions,
recovery outcome,
fleet-level audit after delayed consolidation.

This is a time-series problem as much as an AI problem. The value is not just text memory or vector recall. The value is replaying action evidence with SQL, ASOF JOINs, windows, suppression audits, and bounded edge-to-fleet transfer.

What the current research shows

1.00 recovery Top-1

In the Physical AI fixture, the context-gated Action-Outcome variant selected the expected recovery action across all five incident families.

1.00 risky-repeat avoidance

Similar-incident, runbook-prior, and reflection-only baselines repeated hazardous actions. The context-gated variant avoided them.

32 suppressions exposed

Misleading historical successes are still retained for audit, but down-weighted before action aggregation.

Native SQL replay passed

The result survives live ZeptoDB materialization through action/outcome JOINs, ASOF joins, ROW_NUMBER, LAG, and spatial checks.

Read the evidence table →

Runtime shape

edge event
  -> robot state / sensor summary
  -> historical action outcomes
  -> context gate
  -> allow, suppress, or manual review
  -> decision ledger
  -> edge outbox
  -> bounded fleet feed
  -> fleet audit and policy learning

ZeptoDB’s current path separates immediate edge behavior from fleet-wide learning:

Edge-local memory makes the immediate safety decision.
Fleet-global memory receives evidence later for audit, replay, and policy improvement.
ACK ledgers make delivery boundaries explicit.
Commit ledgers make supervisor decisions idempotent across replay and restart.
Admin-gated runtimes keep pilot scope explicit instead of pretending this is default-on autonomy.

What ZeptoDB provides

Temporal evidence

Store robot incidents, state, sensor summaries, actions, retrieval evidence, suppressions, outcomes, and ACK rows as queryable time-series tables.

SQL replay

Use ASOF JOIN, action/outcome JOIN, suppression audit JOIN, ROW_NUMBER, LAG, and spatial predicates to replay why an action was allowed or suppressed.

Bounded edge-to-fleet feed

Transfer decision, retrieval, and suppression evidence through bounded batches with duplicate, late, outage, and restart handling.

Shadow supervisor

Test recommendations in shadow mode, suppress hazardous proposals, preserve manual-review decisions, and avoid duplicate work after restart.

Maturity boundary

This is an active product track, not a loose idea. It is also not a broad GA control feature today.

Area	Current status
Offline Physical AI comparison	Research complete
Native ZeptoDB SQL replay	Research complete
Edge/fleet bounded feed semantics	Research complete
C++ connector and SQL/HTTP runtime path	Experimental runtime path
Action-Outcome shadow supervisor	Experimental runtime path
Edge/fleet SQL/HTTP adapter	Controlled pilot scope
Default-on operator feature / GA autonomous control	Not promoted

The current supported language is: controlled pilot / controlled shadow pilot, with explicit non-goals, monitoring, rollback, and promotion gates.

Where it applies

Warehouse robots

Avoid repeating risky route decisions after slip, occlusion, geofence, or localization incidents.

Robot arms

Suppress repeated torque-limit increases when similar recoveries required pause, recalibration, or speed reduction.

Drones and mobile robots

Compare continue-mission, return-to-base, clean-sensor, and navigation-mode actions against previous outcomes.

Cold-chain and logistics

Preserve action, environmental context, and recovery outcome for fleet-level audit after edge decisions.

Next steps

Research Evidence Experiment 013-023 results and maturity boundaries

Physical AI & Robotics ROS 2 ingest, typed profiles, edge telemetry, and replayable robot episodes

Agent Memory Live evidence, scoped recall, prompt cache, and AgentOps replay

Research Overview Grounded Physical AI memory scope, limits, and research-only claims