Mastra Changelog 2025-01-31

Welcome to Mastra Changelog #4. This release introduces persistent storage with the Memory API, replaces Docker/Postgres with LibSQL, and splits vector stores into modular packages. The core package is now 50% smaller with improved tree-shaking.

And....stick around to the end for the "one more thing" drop.

Memory API

The new Memory API adds persistent storage and vector search capabilities:

• MastraStorage and MastraVector classes for unified memory management
• Support for both PostgreSQL and Redis as backend storage options
• Token window management with automatic date-based filtering
• Memory management helpers in stream and generate:
  • lastMessages: Control number of recent messages included
  • semanticSearch: Configure semantic search with:
    • topK: Number of similar messages to retrieve
    • messageRange: Messages before/after each result

LibSQL replacing @mastra/engine

We've replaced @mastra/engine with LibSQL as our storage layer, including:

• Persistent storage for agent messages and conversation history
• Tool call tracking and persistence
• Workflow state persistence for suspended workflows

The storage layer powers the mastra dev server while eliminating the need for a Docker/Postgres setup locally.

Speech Provider Modules

We've replaced @mastra/tts with a modular speech provider system:

• Azure Speech (@mastra/speech-azure)
• DeepGram (@mastra/speech-deepgram)
• ElevenLabs (@mastra/speech-elevenlabs)
• Play.ai (@mastra/speech-playai)

The play.ai integration is new.

Vector Store Integrations

We've moved our vector store implementations into separate packages for better modularity:

• Cloudflare Vectorize (@mastra/vector-cloudflare)
• LibSQL (@mastra/vector-libsql)
• Astra DB (@mastra/vector-astra)
• Chroma (@mastra/vector-chroma)
• Pinecone (@mastra/vector-pinecone)
• Qdrant (@mastra/vector-qdrant)
• Upstash (@mastra/vector-upstash)

Note that the Vectorize and Libsql implementations are new ones! Also, the pg filter implementation has been updated, and we've split embed into separate embed and embedMany functions for more consistent return types.

The vector store implementations each (still) include:

• Standard methods: upsert, query, createIndex, listIndexes
• Configurable similarity metrics (cosine, euclidean, dotproduct)
• Metadata filtering support
• Automatic UUID generation for vector IDs

Unified Filter API

The new MongoDB-style filter API is based on sift, providing a standard filtering interface across vector stores:

• Query operators: $eq, $ne, $gt, $gte, $lt, $lte
• Logical operators: $and, $or, $not, $nor
• Array operators: $in, $nin, $all, $elemMatch
• Element operators: $exists
• Regex operators: $regex, $options

The API enables hybrid RAG search by combining vector similarity with metadata filtering:

const results = await vectorStore.query({
  vector,
  filter: {
    date: { $gt: "2024-01-01" },
    tags: { $in: ["documentation", "api"] },
    $or: [{ status: "published" }, { author: "admin" }],
  },
});

Each vector store translates these filters to its native query language, maintaining consistent behavior across implementations.

Core architecture improvements

Core architecture improvements:

• Split core package into separate entry files for better tree-shaking
• Updated Agent and LLM configuration with granular temperature controls
• Implemented unified bundling architecture across:
  • Mastra deployers with optimized asset handling
  • Vector store modules with type-safe interfaces
  • Logging system with structured output

The overall @mastra/core package is now 50% smaller and the actual code imported will be even smaller than that (due to better tree-shaking).

Documentation Updates and bug fixes

• Added intermediate RAG examples with code samples
• Updated vector store integration guides with configuration examples
• Added package keyword tags for better discoverability
• Updated package READMEs with usage patterns
• Added comprehensive RAG testing suite
• Fixed workflow condition display in graph visualization
• Resolved dev watcher file change detection issues
• Fixed 500 error in memory persistence operations
• Fixed thread sidebar state management in playground agent chat
• Fixed logger configuration conflicts in deployment
• Fixed CLI create command component flag parsing

....and one more thing

We built Mastra Cloud as a serverless platform for AI apps

The first app we deployed on it was this pretty cool beat generator

Stay tuned for more :)