TL;DR

When a Mastra Memory uses MongoDBVector with Automated Embedding, semantic recall should let MongoDB embed messages and queries server-side instead of requiring a client-side embedder. Today the docs and code both require an embedder, and the SemanticRecall processor always embeds client-side and passes a numeric queryVector, so it can't benefit from server-side embedding. Depends on NODE-7734.

Context

What semantic recall is. Mastra's Memory can retrieve past messages by meaning ("semantic recall") rather than recency. Each saved message gets an embedding (a numeric vector) stored in a vector store; at query time the incoming text is embedded and used to find nearby vectors. See the live docs: https://mastra.ai/docs/memory/semantic-recall.

How it works today. Both the docs and the code require a client-side embedder:

• The docs Quickstart states you must "provide a vector store and embedder," and every example (including the MongoDB tab) passes one.
• The SemanticRecall processor (packages/core/src/processors/memory/semantic-recall.ts) does the embedding itself on both sides: on save it embeds each message client-side and upserts the resulting vectors; on recall it embeds the query client-side and calls the store's {{query( { queryVector }

  )}}. It reaches the store through the generic MastraVector interface, so this is not MongoDB-specific plumbing.

• The public recall() method (with vectorSearchString) is a second entry point that flows through the same processor.

The problem. NODE-7734 adds MongoDB Automated Embedding, where the database embeds text server-side (no client-side embedder needed). But Memory still requires an embedder when semanticRecall is enabled and always passes a precomputed queryVector. A MongoDB user who wants server-side embedding is forced to also configure and pay for a client-side embedder that does redundant work.

Desired User Experience

Before — client-side embedder required (current, documented):

import { Memory } from '@mastra/memory';
import { MongoDBStore, MongoDBVector } from '@mastra/mongodb';
import { ModelRouterEmbeddingModel } from '@mastra/core/llm';

const memory = new Memory({
  storage: new MongoDBStore({ id: 'mem', uri: process.env.MONGODB_URI!, dbName: 'app' }),
  vector: new MongoDBVector({ id: 'vec', uri: process.env.MONGODB_URI!, dbName: 'app' }),
  embedder: new ModelRouterEmbeddingModel('openai/text-embedding-3-small'), // required today
  options: { semanticRecall: { topK: 3, messageRange: 2 } },
});

After — MongoDB embeds server-side (desired):

import { Memory } from '@mastra/memory';
import { MongoDBStore, MongoDBVector } from '@mastra/mongodb';

const memory = new Memory({
  storage: new MongoDBStore({ id: 'mem', uri: process.env.MONGODB_URI!, dbName: 'app' }),
  // Store configured for Automated Embedding (per NODE-7734) — no client-side embedder needed.
  vector: new MongoDBVector({ id: 'vec', uri: process.env.MONGODB_URI!, dbName: 'app' }),
  options: { semanticRecall: { topK: 3, messageRange: 2 } }, // no `embedder`
});

// The explicit recall() entry point works the same way — MongoDB embeds the search string server-side.
const { messages } = await memory.recall({
  threadId: 'thread-123',
  vectorSearchString: 'What did we decide about the deadline?',
  threadConfig: { semanticRecall: true },
});

Task

Goals

• Let Memory semantic recall work with a server-embedding store, without a client-side embedder.
• Keep the existing client-side embedder path working unchanged for all other stores.

Where

• Recall processor: packages/core/src/processors/memory/semantic-recall.ts
• Memory config/validation: packages/core/src/memory/memory.ts
• Docs to update (live): https://mastra.ai/docs/memory/semantic-recall — source docs/.../memory/semantic-recall.mdx (Quickstart MongoDB tab + "Embedder configuration" section)

Acceptance Criteria

• [ ] A user can enable semanticRecall with MongoDBVector (Automated Embedding) without configuring an embedder, and recall works end-to-end — both via an agent and via the explicit recall( { vectorSearchString }) method.
• [ ] The existing client-side embedder path is unchanged for all stores that don't embed server-side; embedder remains required for those.
• [ ] Memory raises a clear, actionable error if semanticRecall is enabled with neither a client-side embedder nor a server-embedding store.
• [ ] Integration tests cover MongoDB Automated Embedding recall (save + recall, both entry points) and confirm existing client-side recall tests still pass.
• [ ] A follow-up documentation ticket is created (or the semantic-recall page is updated) to add a no-embedder MongoDB configuration and adjust the "embedder is required" framing.

Additional Context

• Depends on NODE-7734: the store must first support text-based upsert and text-based query.
• The docs' latency rationale for disabling recall ("each call converts messages into embeddings") is reduced on the server-side path, since embedding no longer costs a client round-trip — the doc update should reflect this.

References

• NODE-7734 (MongoDB Automated Embedding in the vector store): https://jira.mongodb.org/browse/NODE-7734