Mastra Changelog 2025-02-10

This week we focused on streamlining Mastra's core architecture through better integration with AI SDK, optimized bundle sizes, and improved developer tooling.

We've made significant changes to model routing, introduced subpath imports for better tree shaking, and added tracing capabilities to our storage layer.

Users should be using ai-sdk for model routing

We removed Mastra's own model routing implementation (which was built on top of the Vercel AI SDK) and instead recommend using ai-sdk directly for model routing. This makes it clear that the happy path for Mastra is using AI SDK and Mastra together.

Here's how to use AI SDK with Mastra:

import { openai } from '@ai-sdk/openai';
import { Agent } from '@mastra/core/agent';

const agent = new Agent({
  name: 'WeatherAgent',
  instructions: 'Instructions for the agent...',
  model: openai('gpt-4-turbo'),  // Model comes directly from AI SDK
});

// Use the agent
const result = await agent.generate('What is the weather like?');

Eventually we will also support alternate model routing providers, like Langchain, but that is not part of this release.

In practice, what this means is that functions like embed and embedMany should be imported directly from ai-sdk, and agents now take models from ai-sdk directly, rather than @mastra/core.

We removed the LLM object on the Mastra class.

New short-term working memory

The short-term working memory system uses XML-like tags to track and update contextual information during conversations. When enabled, agents can maintain persistent information through a stream-based approach where memory updates are automatically managed through tagged data in the conversation stream, with the option to mask these updates from end users.

New Bundling System with Subpath Imports

Mastra has added a bundling system using Rollup with tree shaking enabled to eliminate unused code across packages.

In order to take advantage of this, we recommend using the recommended import syntax:

import { Agent } from '@mastra/core/agent';

This is a change from the previous pattern of import { Agent } from '@mastra/core';.

Along with the recent model routing changes, the core bundle size is down to 270kb from around 1MB.

DefaultStorage primitive with LibSQL

Mastra now uses LibSQL as the default storage primitive. Here's how to use it:

import { DefaultStorage } from '@mastra/core/storage';

const storage = new DefaultStorage({
  config: {
    url: ':memory:',
  }
});

// Initialize tables
await storage.init();

The storage primitive handles:

• Thread and message storage
• Workflow snapshots
• Evaluation results

Mastra storage includes tracing (soon)

Mastra's LibSQL storage layer includes tracing, and we're working on adding a UI in mastra dev for it. We should merge this in the next couple of days: https://github.com/mastra-ai/mastra/pull/1820