MongoDB announced on November 5, 2026 that its integration for the Vercel AI SDK is now generally available as an open source npm package. The package lets TypeScript applications and agents built on Vercel's toolkit keep their memory in MongoDB Atlas.
The Vercel AI SDK covers model calls, streaming, tools and multi-step agents, but it has no database of its own. That means an app forgets the user and the previous conversation as soon as a run ends. The new package fills that gap without asking teams to design their own memory layer.
What MongoDB released
The integration ships as @mongodb-developer/vercel-ai-memory on npm under the Apache 2.0 license. Once installed, it connects an AI SDK project to an Atlas cluster and writes what the app learns into documents. Those memories survive new sessions, different users and process restarts.
According to MongoDB, building this by hand usually means a schema for memories, cleanup jobs for stale data and a vector index for meaning-based lookups. The package handles those pieces for you. It works with AI SDK versions 6 and 7 and needs an Atlas cluster.
This is a separate release from the Google ADK integration MongoDB shipped earlier, which we covered in our story on MongoDB's Google ADK integration for agent memory. The new package targets the JavaScript and TypeScript world around Vercel's SDK.
Five memory types with their own retention
The package sorts every memory into one of five categories and stores each in Atlas. Each category has a different lifetime and is looked up in a different way:
- Session: the running conversation, which keeps the latest messages in view.
- Semantic: facts and preferences about people and entities, so users do not repeat themselves.
- Procedural: workflows and instructions that shape how the app performs tasks.
- Episodic: important events and their outcomes, kept by default until a year of inactivity.
- Scratchpad: short-lived working notes that expire after one hour, with an option to promote them to episodic memory.
Semantic and procedural entries are versioned, so older values remain in history after an update. Lookups in those two tiers, plus episodic memory, run through MongoDB Vector Search, which matches by meaning rather than exact words.
At startup the package builds the needed vector indexes and probes the embedding size on its own. Any embedding model supported by the AI SDK can be used, and MongoDB names Voyage AI, OpenAI, Cohere and Google as examples. The GitHub README also describes an optional Atlas Automated Embedding mode that creates vectors on the server.
How forgetting and production controls work
Each tier comes with default retention settings that developers can change. Four retention modes are available: none, which never expires; ttl, which expires after a fixed time; ttl+importance, which keeps items flagged as important; and dynamic, where you set your own forgetting curve.
A tool called memory_forget lets the model remove a specific entry when a user asks it to drop something. MongoDB says the deletion normally completes in roughly a minute. Because data, memory and vector search all live in one Atlas database, there is no separate vector store to secure and scale.
For production use, MongoDB highlights several controls:
- Deterministic session history: by default the model chooses when to save and read turns, which can miss some. AI SDK hooks can instead load and save the full transcript on every turn, storing each user, assistant and tool message once.
- Tenant isolation: adding a field like
tenant_idto the vector indexes keeps each search inside one customer's data. - Naming and layout: collections and indexes can be renamed to fit an existing schema, and unused tiers can be switched off.
- Model freedom: the package works with any model the AI SDK supports.
What it means for developers
For teams already on the Vercel AI SDK, this removes a lot of plumbing. Instead of building a memory service, you add one package, point it at your Atlas cluster and choose an embedder. If your app data already sits in MongoDB, keeping memory beside it simplifies operations.
The session mode deserves a careful choice. The README calls the tool-driven mode fine for demos but not deterministic, and it recommends the hook-driven mode for production. If you upgrade from AI SDK 6 to 7, note that the hook context key changed, and package version 0.6.0 handles both names.
Multi-tenant SaaS builders should add a tenant filter field from day one rather than retrofitting it later. Also review the default retention per tier, since keeping everything raises cost and slows retrieval, while short limits make the app forget useful context. The README marks the separate Vercel Eve provider as not yet production-ready.
If you are still mapping out the AI skills these projects need, our artificial intelligence roadmap for developers is a good starting point.
How to get started
MongoDB describes three steps. First install the package along with its peer dependencies, which the README lists as ai, mongodb and zod. Next create a memory instance with your Atlas URI and an embedding model, then pass the memory tools to your agent scoped to a user ID and session ID.
From that point the model decides when to store and search memories. The GitHub repository contains the source and runnable examples, including a deterministic agent that uses the hook-driven session mode.
Conclusion
The MongoDB Vercel AI SDK integration gives TypeScript AI apps a ready-made, structured memory layer in Atlas with vector search, retention policies and tenant filtering included. It is open source, generally available and supports AI SDK 6 and 7, so teams can try it in an existing project with little setup.
Sources
- Introducing the MongoDB Vercel AI SDK Integration: Persistent Memory for AI Applications (MongoDB Blog)
- mongodb-developer/vercel-ai-memory (GitHub)
- @mongodb-developer/vercel-ai-memory (npm)
See my articles more often when you search Google.
