Persistence Stores

Builds on ai-persistence and ai-persistence/server.

@tanstack/ai-persistence ships contracts, not a backend for your database. An adapter is an object with a stores map; implement the stores you need against whatever you already run and hand the result to withPersistence. The core never inspects your tables, so the schema is yours.

Use memoryPersistence() for dev and tests. Everything durable is an adapter you write. This skill is the contract reference; the per-stack recipes that write a chat-persistence.ts into an app are ai-persistence/build-{drizzle,prisma,cloudflare,custom}-adapter, and a complete node:sqlite implementation lives in examples/ts-react-chat/src/lib/sqlite-persistence.ts.

Choose a shape

import { defineAIPersistence } from '@tanstack/ai-persistence'
import type { ChatWithInterruptsPersistence } from '@tanstack/ai-persistence'

// Sparse is fine — only implement what you need.
export const persistence: ChatWithInterruptsPersistence = defineAIPersistence({
  stores: {
    messages, // required for withPersistence / reconstructChat
    runs, // required if you have interrupts
    interrupts,
    // metadata optional
  },
})

defineAIPersistence preserves exact keys and rejects unknown keys at runtime.

Annotate your factory with a named shape. Bare AIPersistence is the all-optional sparse bag, so withPersistence and reconstructChat reject it (stores.messages is possibly undefined). This is the single most common mistake when writing an adapter.

stores accepts exactly four keys — messages, runs, interrupts, metadata. Anything else (notably locks or sandbox instance maps) throws Unknown AIPersistence store key at runtime and fails to type-check. Locks: ai-core/locks / @tanstack/ai/locks. Sandbox instance resume: @tanstack/ai-sandbox.

Contracts and invariants

MessageStore

interface MessageStore {
  loadThread(threadId: string): Promise<Array<ModelMessage>>
  saveThread(threadId: string, messages: Array<ModelMessage>): Promise<void>
}

• loadThread → [] for unknown threads (never null).
• saveThread is a full overwrite, not append. A one-message payload wipes history.

RunStore

RunStatus, TerminalRunStatus, RunRecord, RunStore, defineRunStore, and isTerminalRunStatus are defined in @tanstack/ai and re-exported from @tanstack/ai-persistence. Import those from either; the recipes in this skill import from @tanstack/ai-persistence so an adapter author needs only one package name.

RunError is the exception — it is NOT re-exported. Import it from @tanstack/ai directly (import type { RunError } from '@tanstack/ai'); the @tanstack/ai-persistence barrel has no such export and the import fails to resolve.

Four methods are required (createOrResume / update / get / findActiveRun). Two are optional: implement only the ones your backend needs, and leave the rest off the object entirely (not undefined, just absent). A four-method RunStore is a fully valid backend.

withPersistence itself calls none of the three non-createOrResume/update query methods, so leaving both optional ones off costs nothing in the middleware. Their consumers are elsewhere, and each absence disables exactly one feature:

Consumers of the two OPTIONAL methods feature-detect with store.method?.(...) and degrade rather than throwing. findActiveRun is required, so nothing feature-detects it.

The conformance testkit does not feature-detect. An optional method that is missing and not declared in skipMethods fails the suite, so an omission is always a choice you made on purpose rather than a check that quietly did not run. Declare yours and the suite reports them as skipped with a reason:

// The shipped sqlite example implements findActiveRun and listReclaimable and
// declares only the one it omits.
runPersistenceConformance('sqlite', () => persistence, {
  skipMethods: ['runs.listByThread'],
})

interface RunStore {
  // Required
  createOrResume(
    input: Pick<RunRecord, 'runId' | 'threadId' | 'startedAt'> & {
      status?: RunStatus
    },
  ): Promise<RunRecord>
  update(
    runId: string,
    patch: Partial<
      Pick<
        RunRecord,
        | 'status'
        | 'finishedAt'
        | 'error'
        | 'usage'
        | 'sandboxKey'
        | 'detachedSince'
        | 'cancelRequested'
        | 'driverEpoch'
      >
    >,
  ): Promise<void>
  get(runId: string): Promise<RunRecord | null>
  findActiveRun(threadId: string): Promise<RunRecord | null>

  // Optional
  listByThread?(threadId: string): Promise<Array<RunRecord>>
  listReclaimable?(opts: {
    now: number
    ttlMs: number
  }): Promise<Array<RunRecord>>
}

RunStatus is 'running' | 'interrupted' | 'completed' | 'failed' | 'aborted'. 'interrupted' is a human-in-the-loop pause, not terminal: it is what interrupt-resume continues from, and must never be conflated with 'aborted' (an explicit cancellation). TerminalRunStatus narrows to 'completed' | 'failed' | 'aborted'. isTerminalRunStatus(status) is a type predicate: (status: RunStatus) => status is TerminalRunStatus, so calling it inside a guard narrows status to TerminalRunStatus for the rest of that branch, with no cast needed.

RunRecord.usage is optional. withPersistence sums reported numeric fields across provider calls for that runId, while opaque providerUsageDetails retains the latest reported bag. Known usage is persisted on interruption and every terminal status.

RunRecord.error is a structured RunError, not a bare string:

interface RunError {
  message: string
  code?: string
}

message is the provider's prose (it changes between model versions and cannot be branched on); code is the stable, machine-branchable classification a consumer switches over to retry, escalate, or show specific UI. Store both, and omit code from a mapped record when its column is null rather than writing code: undefined (...(row.errorCode != null ? { code: row.errorCode } : {})).

defineRunStore<const T extends RunStore>(store: T): T returns the passed object's own type, so an optional method your store implements (say, listByThread) stays known-present on the returned value instead of widening back to RunStore's | undefined. You get autocomplete and contract checking without a separate : RunStore annotation, and without a feature-detection guard on your own return value.

The durable-agent-runs fields: sandboxKey, detachedSince, cancelRequested, driverEpoch

These four RunRecord fields exist for the sandbox/durable-run layer to reattach a run a client disconnected from, and for out-of-band cancellation. A RunStore you write must round-trip all four through update → get, even if your app does not use sandboxes yet — the conformance testkit checks this unconditionally (it is not behind skipMethods, because update/get are REQUIRED methods).

• sandboxKey — compound key identifying the sandbox this run is bound to, so a reclaimer can find it to tear down.
• detachedSince — epoch ms when the last viewer detached; absent while someone is attached. Read by listReclaimable.
• cancelRequested — set by an explicit out-of-band cancel (see below), distinct from a mere disconnect.
• driverEpoch — monotonic fencing token, bumped by each host that claims the run, so a superseded host can discover it lost by comparing the stored value against the one it holds.

Fresh-run reads must be undefined, not a coerced falsy default. A backend that reads a NULL/absent column back as cancelRequested: false or driverEpoch: 0 is claiming knowledge it does not have ("explicitly not cancelled") — that is a different fact from "never set". Omit the field from the mapped record instead (...(row.cancel_requested != null ? { cancelRequested: row.cancel_requested !== 0 } : {})).

update must use 'field' in patch, not patch.field !== undefined, for these four. A caller clears detachedSince on reattach by passing it explicitly as undefined — store.update(runId, { detachedSince: undefined }) — and that must write NULL, not be filtered out of the write. Checking !== undefined cannot tell "clear this field" apart from "I didn't mention this field", so it silently drops the clear and the run looks permanently detached to the reaper forever after. 'detachedSince' in patch is true for an explicit undefined and false when the caller omitted the key entirely — that is the distinction you need. The same applies to cancelRequested (false is a real, meaningful value, not "unset") and to sandboxKey / driverEpoch. See examples/ts-react-chat/src/lib/sqlite-persistence.ts for a worked implementation of exactly this pattern.

Out-of-band cancellation

Cancel intent is never inferred from a disconnect — a user pressing Stop and a user closing the tab produce an identical connection close, so the two are indistinguishable from the abort alone. @tanstack/ai exports the actual primitives:

• requestRunCancel — records durable cancel intent (writes cancelRequested: true through a RunStore), for a run being driven on a host other than the one handling the cancel request.
• wasCancelRequested — reads that intent back.
• RUN_CANCEL_REASON — the well-known abort reason string used for the in-process case (the same host aborting its own signal), paired with isCancelRequestedReason to check for it.

A RunStore you write does not call these directly — they operate on your store through update/get — but cancelRequested must round-trip faithfully (previous section) for the durable path to work at all.

• createOrResume (required): if runId exists, return it unchanged, including its stored usage, and ignore the passed threadId / startedAt / status. Resuming a run does not reset startedAt or overwrite its current status. Idempotent retries and double-submit depend on this. status defaults to 'running' on first creation.
• update (required): missing runId is a no-op (do not throw, do not insert).
• get (required): current record, or null when unknown.
• listByThread (optional): every run for threadId, ascending by startedAt. Only needed to render a thread's past agent activity.
• listReclaimable (optional): runs where status === 'running' AND detachedSince is set AND detachedSince <= now - ttlMs. The cutoff is inclusive: a run detached exactly at the cutoff qualifies. This is a query, not automatic behavior — the consumer is reapDetachedRuns from @tanstack/ai-sandbox, which the application schedules itself (cron, queue, alarm(), waitUntil), so returning this list has no side effect until that sweep runs. detachedSince is written for you by withSandbox's detach path (alongside sandboxKey) and cleared by the takeover path; drop either field and nothing can reclaim the sandbox. cancelRequested is written by requestRunCancel and read by wasCancelRequested, and the reaper's expiry path goes through requestRunCancel to stop a run past its TTL.
• findActiveRun (required): the most recent 'running' run for threadId (max startedAt), or null if none is active. Enables reconnect from a stable thread id without a client-held run id. Stub it out and reconnect silently stops working — null is also the correct answer for an idle thread, so nothing can detect the difference. It was optional for exactly one release cycle and cost precisely that, which is why it is required now.

Capability tiers belong at the STORE level (omit runs entirely and declare ChatTranscriptStores), not the method level — never ship a RunStore with a stubbed method. The two list queries above are the only method-level options, and each must be declared via skipMethods when absent.

InterruptStore

interface InterruptStore {
  create(record: Omit<InterruptRecord, 'status' | 'resolvedAt'>): Promise<void>
  resolve(interruptId: string, response?: unknown): Promise<void>
  cancel(interruptId: string): Promise<void>
  get(interruptId: string): Promise<InterruptRecord | null>
  list(threadId: string): Promise<Array<InterruptRecord>>
  listPending(threadId: string): Promise<Array<InterruptRecord>>
  listByRun(runId: string): Promise<Array<InterruptRecord>>
  listPendingByRun(runId: string): Promise<Array<InterruptRecord>>
}

• create always births 'pending'; insert-if-absent on interruptId (never clobber resolved back to pending).
• All list* ordered by requestedAt ascending.
• Requires a runs store when used with chat persistence.

MetadataStore

interface MetadataStore {
  get(namespace: string, key: string): Promise<unknown | null>
  set(namespace: string, key: string, value: unknown): Promise<void>
  delete(namespace: string, key: string): Promise<void>
}

• The first argument is an app-defined namespace string, not the Scope identity type — despite SQL backends conventionally naming the column scope.
• Identity is two fields (namespace, key) — do not join with : (('a:b','c') and ('a','b:c') must stay distinct).
• Stored null is type-indistinguishable from absence; wrap if you must persist real null ({ value: null }).
• SQL backends usually reject nullish set (NOT NULL JSON columns) with a clear TypeError — match that or document your semantics.

Timestamp convention

Store records (RunRecord, InterruptRecord) speak epoch milliseconds (number). Wire/result references that leave the persistence layer speak ISO-8601 strings; the middleware converts at the boundary. Do not mix the two on one field.

Minimal message store example

Type each store with its define*Store helper (defineMessageStore, defineRunStore, defineInterruptStore, defineMetadataStore): pass the object literal and get autocomplete + contract checking inline, with no : MessageStore annotation. The result composes into defineAIPersistence with exact presence.

import { defineMessageStore } from '@tanstack/ai-persistence'
import type { ModelMessage } from '@tanstack/ai'

const threads = new Map<string, Array<ModelMessage>>()

export const messages = defineMessageStore({
  async loadThread(threadId) {
    return [...(threads.get(threadId) ?? [])]
  },
  async saveThread(threadId, next) {
    threads.set(threadId, [...next])
  },
})

For durable DBs, preserve the same semantics with upserts / full-row replace.

Adopt part of it

You rarely need all four stores in the same system. Implement the ones you own and fill the rest from another base with composePersistence:

import { composePersistence, memoryPersistence } from '@tanstack/ai-persistence'
import { messages, runs } from './my-postgres-stores'

export const persistence = composePersistence(memoryPersistence(), {
  overrides: { messages, runs },
})

Only listed keys move; others stay on the base. Pass false to drop a store. There is no cross-store transaction — if messages lives in Postgres and interrupts in Redis, a write touching both is two writes. The store invariants (idempotent createOrResume, insert-if-absent create) are exactly what make those retries safe.

composePersistence accepts the four state keys. Locks and sandbox instance maps are not composable here.

Map onto an existing schema

• Your column names, your types. Name columns anything; use jsonb, timestamptz, whatever — convert in the row mapper. The record shape the methods return is fixed; how you store it is not.
• Extra columns are fine. Add user_id, audit columns, a tenant id. Keep them nullable or defaulted so the store's inserts still succeed. The stores never read or write columns they do not know about.
• Omit absent optionals in row mappers (...(row.error != null ? { error: row.error } : {})) so records compare cleanly.

Authorization

Store methods take bare threadIds. Authorize at the route before loadThread / saveThread / reconstructChat({ authorize }). Derive user identity from session, not the client body alone.

Conformance tests (required)

import { runPersistenceConformance } from '@tanstack/ai-persistence/testkit'
import { myPersistence } from '../src/persistence'

runPersistenceConformance('my-backend', () => myPersistence())

// Declare intentional omissions. The suite covers all seven stores, so a
// chat-only backend skips the generation half:
// runPersistenceConformance('chat-only', () => p, {
//   skip: ['generationRuns', 'artifacts', 'blobs'],
// })
// `skip` never accepts 'locks' — locks are not a store.

// Declare an intentionally-unimplemented OPTIONAL RunStore method with
// skipMethods, so vitest reports it as a real SKIPPED case:
// runPersistenceConformance('my-backend', () => myPersistence(), {
//   skipMethods: ['runs.listByThread', 'runs.listReclaimable'],
// })

The testkit is the compatibility gate: round-trips, rich message shapes, empty-thread [], createOrResume idempotency, interrupt insert-if-absent, list ordering, composite-key non-aliasing. A missing store that is not listed in skip fails loudly.

skip accepts only 'messages' | 'runs' | 'interrupts' | 'metadata'. Do not pass 'locks' — it is not a state store and the suite does not cover it.

skipMethods (declare-or-fail for optional RunStore methods). A backend that omits an OPTIONAL RunStore method (listByThread, listReclaimable — findActiveRun is required and cannot be declared away) must declare it in skipMethods as 'runs.<method>', e.g. skipMethods: ['runs.listByThread', 'runs.listReclaimable']. An omitted method that is NOT declared throws with an actionable message instead of silently reporting a pass; a declared one is reported as a SKIPPED vitest case, never as a pass. A case that did not run must never be indistinguishable from one that did. See examples/ts-react-chat/src/lib/sqlite-persistence.test.ts for a worked example: it declares skipMethods: ['runs.listByThread'] only, keeping both findActiveRun and listReclaimable under test.

Reference implementation: memoryPersistence() in @tanstack/ai-persistence.

Common mistakes

CRITICAL: Append-only saveThread

Breaks the authoritative-history contract.

CRITICAL: createOrResume overwriting existing runs

Breaks safe resume / double-submit.

CRITICAL: Interrupt create upserting to pending

Can resurrect a resolved approval.

HIGH: Returning bare AIPersistence from the factory

withPersistence rejects it. Annotate a named shape.

HIGH: list* without stable requestedAt order

Middleware and tests assume ascending order.

HIGH: Skipping the testkit

Silent semantic drift shows up as stuck approvals or wiped history in prod.

HIGH: listReclaimable cutoff off by one

The cutoff is inclusive (detachedSince <= now - ttlMs). Using a strict < drops runs detached exactly at the boundary.

HIGH: Treating listReclaimable as automatic reclamation

It is a query a caller runs, not something the package acts on by itself. Nothing reaps a returned run for you.

Cross-references

• ai-persistence/server — when middleware calls each store
• ai-persistence/build-drizzle-adapter / -prisma- / -cloudflare- / -custom- — per-stack recipes
• ai-core/locks — not a state store