@tanstack/ai-persistence weekly downloads statistics

@tanstack/ai-persistence

Composable state persistence for TanStack AI messages, runs, interrupts, metadata, and locks

A conversation that only lives in memory is gone on reload and absent on a second device. This package stores it in your own database: one middleware on the server writes the transcript — and, with the matching stores configured, run status and pending approvals — through an adapter you define, and the client asks the server for the thread on mount. The client half ships in the framework package you already use (@tanstack/ai-react, -vue, -solid, -svelte, -angular, or @tanstack/ai-client).

Installation

Usage

Server: store the conversation

withPersistence writes the transcript into your messages store, plus run status and pending approvals when the adapter also provides runs and interrupts. Start with memoryPersistence() for local development and swap in your own adapter later:

Client: bring it back

persistence: true puts the server in charge: the browser caches nothing and fetches the thread on mount. Best for multi-user and multi-device apps.

Prefer the browser to own the history? Pass a storage adapter instead — localStoragePersistence(), sessionStoragePersistence(), or indexedDBPersistence() — and no server store is needed.

Survive a reload mid-answer

Add a GET on the same route. reconstructChat returns the stored thread plus a cursor to any run still generating; useChat tails that run so the reply finishes in place.

Build your own adapter

An adapter is a plain object of store functions. The core never looks at your tables, so the schema stays yours. One store, messages, is enough for withPersistence:

The other stores are optional and add capabilities as you provide them: runs, interrupts, and metadata for chat state; generationRuns, artifacts, and blobs for image, video, speech, and transcription runs. composePersistence layers overrides on top of a base backend.

Prove it with the conformance suite

The same suite every packaged backend runs is shipped for yours:

Stores you do not provide go in skip. The testkit is a Vitest suite; vitest is an optional peer dependency, so install it in your project before importing @tanstack/ai-persistence/testkit.

Documentation

• Overview: the three steps, and which setup you want • Chat persistence: the server middleware in full, including durable interrupts • Client persistence: the modes, storage backends, and what a reload restores • Build your own adapter • Controls: compose backends per store • Store reference: every store's methods and how the records relate

License

MIT