Persistence Controls#

Persistence has no feature flags. What you persist is decided by which state stores the backend provides, and you compose backends per store. Supply only the stores your workflow needs.

Need a mutex across instances? See Locks below.

Named shapes (prefer these)#

There is no public sparse AIPersistenceStores export — use a named shape or AIPersistence<{ messages: MessageStore, … }> for custom maps. defineAIPersistence / composePersistence still accept sparse maps by inference.

What each state store gives you#

withPersistence(persistence) inspects the stores that are present. Store presence is the capability selection mechanism for optional chat features.

Entrypoint requirements#

Compose and override stores#

composePersistence takes the base backend first and an overrides object second. Here it starts from the in-memory reference backend and swaps in custom interrupts / runs stores:

import { composePersistence, memoryPersistence } from '@tanstack/ai-persistence'
// Your own store implementations of the InterruptStore / RunStore contracts.
import { interruptStore, runStore } from './stores'

const persistence = composePersistence(memoryPersistence(), {
  overrides: {
    interrupts: interruptStore,
    runs: runStore,
  },
})

Each override is independent:

import { composePersistence, memoryPersistence } from '@tanstack/ai-persistence'

// Drop metadata; the resulting type has no `metadata` key.
const withoutMetadata = composePersistence(memoryPersistence(), {
  overrides: { metadata: false },
})

Unknown store names fail type checking, and are also rejected at runtime when values arrive from untyped JavaScript.

Valid store combinations#

• withPersistence requires messages.
• interrupts requires runs: an interrupt record is scoped to a run.
• withGenerationPersistence requires runs.

To define a partial backend directly rather than by composing, use defineAIPersistence({ stores: { ... } }) and pass only the stores you have. See the store interface reference for the store contracts.

Locks (coordination)#

Locks coordinate work across instances (a distributed mutex). They live in @tanstack/ai/locks and apply as their own middleware with withLocks, alongside withPersistence. Full guide: Locks.

import { withLocks, InMemoryLockStore } from '@tanstack/ai/locks'
import { withPersistence, memoryPersistence } from '@tanstack/ai-persistence'

const middleware = [
  withPersistence(memoryPersistence()),
  withLocks(new InMemoryLockStore()), // multi-instance: distributed LockStore
]