Runtime The Runtime interface executes steps, manages context, and coordinates memory and channels.
import { InMemoryRuntime, step } from '@noetic/core' ; const runtime = new InMemoryRuntime (); const ctx = runtime. createContext ({ state: { count: 0 } }); const add = step. run ( 'add' , async ( n : number , ctx ) => { return n + 1 ; }); const result = await runtime. execute (add, 5 , ctx); // result === 6 The Runtime is the execution engine at the heart of Noetic. It creates contexts, dispatches steps, manages channels and memory layers, and provides tracing infrastructure. You can use the built-in InMemoryRuntime or implement the Runtime interface for durable, distributed, or custom execution environments.
interface Runtime { execute < I , O >( step : Step < I , O >, input : I , ctx : Context ) : Promise < O >; createContext ( opts ?: CreateContextOpts ) : Context ; // Channels send < T >( channel : Channel < T >, value : T , ctx : Context ) : void ; recv < T >( channel : Channel < T >, ctx : Context , opts ?: { timeout ?: number }) : Promise < T >; tryRecv < T >( channel : Channel < T >, ctx : Context ) : T | null ; getChannelHandle < T >( channel : ExternalChannel < T >, executionId : string ) : ChannelHandle < T >; // Memory initLayers ( layers : MemoryLayer [], ctx : Context , storage : StorageAdapter ) : Promise < void >; recallLayers ( layers : MemoryLayer [], input : string , ctx : Context ) : Promise < RecallLayerOutput []>; storeLayers ( layers : MemoryLayer [], response : LLMResponse , ctx : Context ) : Promise < void >; disposeLayers ( layers : MemoryLayer [], ctx : Context ) : Promise < void >; assembleView ( agent : AgentConfig , input : string , ctx : Context ) : Promise < Item []>; // Durability checkpoint ( ctx : Context ) : Promise < void >; restore ( executionId : string ) : Promise < Context | null >; cancel ( ctx : Context , reason ?: string ) : Promise < void >; // Tracing createSpan ( name : string , parent : Span | null ) : Span ; } Method Description execute(step, input, ctx)Run a single step with the given input and context. Returns the step output. createContext(opts?)Create a new execution context. Options let you set parent, initial items, state, threadId, and resourceId.
Option Type Description parentContextParent context (for spawn). itemsItem[]Pre-seed the item log. stateunknownInitial mutable state. threadIdstringConversation thread identifier. resourceIdstringOptional resource/tenant identifier.
Method Description send(channel, value, ctx)Push a value into a channel. recv(channel, ctx, opts?)Wait for a value from a channel. Supports timeout. tryRecv(channel, ctx)Non-blocking read. Returns null if nothing is available. getChannelHandle(channel, executionId)Get a ChannelHandle
The runtime coordinates memory layer lifecycle hooks:
Method Description initLayersInitialize layers at the start of an agent run. Loads persisted state from storage. recallLayersQuery layers for relevant items given an input string. Returns RecallLayerOutput[]. storeLayersWrite new items from an LLM response into layers. disposeLayersClean up layers at the end of a run. assembleViewCombine item log and recalled memory into the final item list sent to the model.
interface RecallLayerOutput { layerId : string ; items : Item []; tokenCount : number ; } Method Description checkpoint(ctx)Persist execution state for crash recovery. restore(executionId)Restore a previously checkpointed context. Returns null if not found. cancel(ctx, reason?)Cancel a running execution.
Method Description createSpan(name, parent)Create a new tracing span . Links to a parent span when provided.
InMemoryRuntime is the built-in implementation. It keeps all state in memory -- no persistence, no external dependencies.
import { InMemoryRuntime, InMemoryExporter } from '@noetic/core' ; const exporter = new InMemoryExporter (); const runtime = new InMemoryRuntime ({ traceExporter: exporter, }); Option Type Default Description callModelCallModelFnundefinedCustom function that calls the LLM. Required for step.llm execution. traceExporterTraceExporterNoopExporterWhere to send completed spans. layerStateStoreLayerStateStoreIn-memory store Backing store for memory layer state.
Key behaviors of InMemoryRuntime:
checkpointrestorenull.assembleView
AgentConfig defines the full configuration for an agent. Patterns like react build one internally, but you can construct it directly for full control.
interface AgentConfig { name : string ; description : string ; model : string ; instructions : string | (() => string | Promise < string >); tools ?: Tool []; outputSchema ?: ZodType ; memory ?: MemoryLayer []; storage ?: StorageAdapter ; projection ?: ProjectionPolicy ; hooks ?: AgentHooks ; } Field Description nameHuman-readable agent name. descriptionWhat the agent does (used in spawn summaries). modelLLM model identifier (e.g., 'gpt-4o'). instructionsSystem prompt. Can be a string or an async function for dynamic prompts. toolsArray of tools the agent can call. outputSchemaZod schema for structured output validation. memoryMemory layers to attach to this agent. storageStorage adapter for memory persistence. projectionPolicy controlling how memory is projected into the context window. hooksLifecycle hooks (see below).
interface AgentHooks { beforeStep ?: ( step : Step , ctx : Context ) => Promise < void >; afterStep ?: ( step : Step , result : unknown , ctx : Context ) => Promise < void >; } Hook When it fires Typical use beforeStepBefore each step executes. Logging, injecting state, guard checks. afterStepAfter each step completes. Metrics collection, state updates, audit logging.
const config : AgentConfig = { name: 'researcher' , description: 'Searches the web and summarizes findings.' , model: 'gpt-4o' , instructions: 'You are a research assistant.' , hooks: { async beforeStep ( step , ctx ) { console. log ( `Starting step ${ step . id } (depth ${ ctx . depth })` ); }, async afterStep ( step , result , ctx ) { console. log ( `Step ${ step . id } used ${ ctx . tokens . total } tokens` ); }, }, };