NOTOYAI
// culture · households · relationships · lore · minds · closure
// PERSISTENT RUNTIME INFRASTRUCTURE FOR AI AGENTS

NOTOYAI

Long-lived agents need more than a model.

NotoyAI gives agent systems structured state, long-horizon operation, causal provenance and reproducible execution.

The runtime owns continuity and committed decisions. Language models can render expression without becoming the source of truth.

The Problem

A good next response is not the same as a persistent agent.

Most agent stacks are optimized around a task, a conversation or the next model call. Long-running systems need more: state that survives, memories with provenance, explicit goals and plans, and decisions that can be inspected and compared.

When an agent operates over days, months or a shared environment, the important questions are not only what happened? They are what did the agent know, what changed, why was this action selected, and can we reproduce it?

More prompting does not create continuity. The missing layer is the runtime around the model.

The Product

Continuity lives between decisions.

NotoyAI separates persistent cognition from probabilistic expression.

Structured state
Typed events, memory, social state, pressures, goals, plans and committed actions remain available beyond one interaction.
Deterministic execution
Stable ordering, bounded action selection, checkpoints and canonical traces make runs inspectable and comparable.
Causal provenance
The runtime records which state and plan supported a committed action. Human-readable explanation is projected from those records rather than invented after the fact.
Expression boundary
Language and animation can render a committed result. They do not own evidence-critical state, select the evidentiary action or rewrite its cause.
// THE MODEL CAN SPEAK. THE RUNTIME MUST REMEMBER.
What Exists Now

Current building blocks, stated at their real scope.

Persistent multi-agent runtime
Agents execute in stable order against shared world state. Committed decisions, events and provenance are recorded for inspection and replay.
Typed runtime contract
An engine-agnostic JSON contract exposes agent state, actions, world subscriptions, environment updates and inspection surfaces without exposing private storage or prompt internals.
Seeded reference-world generation
The current village generator produces cast profiles, roles, households, relationship priors, lore, pressures and engine-binding manifests from a seed. Agent-location closure is checked before export.
Unreal Engine reference integration
The current Unreal integration scans NotoyAI semantic components into an engine-agnostic environment map, validates and registers that map, sends agent locations, and receives runtime state and animation intent.

Direct Unreal PCG ingestion and marketplace-ready production packaging remain roadmap work.

Procedural Intelligence Generation

PCG builds the physical world. PIG builds the cognitive one.

Procedural Intelligence Generation is one application of the runtime: structured populations generated from a seed and executed against persistent world state.

In the current reference environment, that includes roles, households, relationship priors, lore, pressures, locations and runtime bindings. It is not a prompt that produces a character biography. It is machine-readable state that the runtime can execute and inspect.

PIG is the product path for generated worlds. The underlying persistent runtime is intended for agent systems beyond game NPCs.

Evidence

Built to be checked.

NotoyAI preserves the run scope, configuration, environment, commands, manifests and hashes needed to inspect a result. Replay claims refer to canonical decision artifacts, not authored vignettes or generated dialogue.

4
Recorded one-year runs, 10 agents each
87,600
Committed decisions per run
0
LLM calls in the recorded decision path

Each run covers 8,760 ticks. Both same-seed replay pairs produce matching decision hashes. Expression is disabled in this evidence set.

These figures describe this evidence set only. They are not claims about unlimited scale, every runtime mode, all hardware or all future integrations.

Research

From stored experience to testable action.

The current research direction is machine-native persistent cognition: turning stored experience into decision-effective state without giving up inspection or reproducibility. The target chain is explicit.

01stored typed event 02gated observation 03typed memory retrieval with receipts 04belief revision 05goal activation 06explicit plan with an active step 07plan-bound committed action 08provenance-faithful explanation

Each stage must become a typed, testable object and survive replay, intervention and invariance checks before it is presented as a general capability. The repository contains implemented substrate and bounded evidence along this path. A complete general persistent cognitive system, broad transfer, metacognition and robust adaptive planning remain research targets.

Where It Applies

For systems that must remain coherent over time.

The current evidence is strongest in long-running multi-agent simulation. The runtime contract itself is engine-agnostic. NotoyAI is being developed for:

Agent evaluation and reproducibility testbeds Multi-agent simulations and training environments Persistent interactive worlds Long-running agent systems with inspectable histories

Validation outside the current reference environments is ongoing.

Company

Developed in Tallinn. Proven in long-running worlds.

NotoyAI is developed by Igikelts OÜ in Tallinn, Estonia.

Ruhnu Jooksik is the first reference environment: a historical Unreal Engine project used to expose failures in memory, continuity, social state and long-horizon simulation.

Ruhnu is where the runtime is exercised. It is not the definition of the company or the product.

RUHNU ISLAND · 1944-1950 · UNREAL ENGINE 5 · IN DEVELOPMENT

Inspect the runtime.

Read the evidence, review the architecture, or contact us about a technical integration or research question.

info@notoyai.com

We will reply from an @notoyai.com or @igikelts.com address.

NOTOYAI · Persistent runtime infrastructure for AI agents
These minds are not toys.
Developed by Igikelts OÜ · Tallinn, Estonia · © 2026 Igikelts OÜ

These minds are not toys
Scroll to begin