Working state

workingState() is a memory block for one typed value. Use it when the agent needs a current plan, current goal, iteration counter, draft state, or a small set of structured findings that should be updated in place.

import { memory, workingState } from '@crux/core/memory'
import { z } from 'zod'

const state = workingState({
  id: 'state',
  schema: z.object({
    goal: z.string(),
    step: z.number(),
    findings: z.array(z.string()).default([]),
  }),
})

const agentMemory = memory({
  id: 'research-agent',
  store,
  namespace: ({ input }) => `thread:${input.threadId}`,
  blocks: [state],
})

The composed memory renders the current state into prompt context. The block also exposes direct methods for application code:

await state.set(
  { goal: 'Compare vector databases', step: 1, findings: [] },
  { store, namespace: 'thread:abc', memoryId: 'research-agent' },
)

await state.patch(
  { step: 2, findings: ['Upstash supports hybrid search'] },
  { store, namespace: 'thread:abc', memoryId: 'research-agent' },
)

const current = await state.get({
  store,
  namespace: 'thread:abc',
  memoryId: 'research-agent',
})

workingState() validates every write with the supplied Zod schema. Invalid writes throw immediately; they are not silently dropped.

When to Use It

Use working state for data that has one current value. If you need a timeline, use episodes(). If you need extracted long-term knowledge, use facts() or procedures(). If the value only matters for one model call, put it in prompt input instead of memory.

Observability

Reads and writes emit block memory events with blockKind: 'working'. Devtools shows this as block memory with a state view, so you can inspect how the scratchpad changes over time.