Embedding Providers

mememory converts text into vector embeddings for semantic search. The embedding provider determines which model generates these vectors. You can use a local model via Ollama (default, private) or a cloud API like OpenAI.

Supported Providers

Provider	Privacy	Cost	Quality	Default Model	Dimensions
Ollama (local)	Full — no data leaves your machine	Free	Good	nomic-embed-text	768
OpenAI	Data sent to OpenAI API	Paid per token	Excellent	text-embedding-3-small	1536
OpenAI-compatible	Varies	Varies	Varies	Configurable	Varies

Ollama (Default)

Ollama runs locally inside Docker. No API keys, no data leaves your machine.

Configuration

# These are the defaults — no env vars needed for standard setup
EMBEDDING_PROVIDER=ollama
OLLAMA_URL=http://localhost:11434

The nomic-embed-text model is automatically pulled when the Docker stack starts. It produces 768-dimensional vectors.

Using a different Ollama model

# Pull a different model first
docker exec mememory-ollama ollama pull mxbai-embed-large

# Then configure
EMBEDDING_PROVIDER=ollama
EMBEDDING_MODEL=mxbai-embed-large

WARNING

Changing the embedding model changes the vector dimensions. You cannot mix vectors of different dimensions in the same database. See Switching Providers below.

OpenAI

Uses the OpenAI embeddings API. Requires an API key and internet connectivity.

Configuration

EMBEDDING_PROVIDER=openai
EMBEDDING_API_KEY=sk-your-api-key-here

The default model is text-embedding-3-small (1536 dimensions). To use a different model:

EMBEDDING_PROVIDER=openai
EMBEDDING_API_KEY=sk-your-api-key-here
EMBEDDING_MODEL=text-embedding-3-large

Model	Dimensions	Cost (per 1M tokens)
text-embedding-3-small	1536	$0.02
text-embedding-3-large	3072	$0.13
text-embedding-ada-002	1536	$0.10

OpenAI-Compatible APIs

Any API that implements the OpenAI embeddings endpoint format works with the openai provider. This includes Azure OpenAI, Mistral, Together AI, and others.

Azure OpenAI

EMBEDDING_PROVIDER=openai
EMBEDDING_URL=https://YOUR-RESOURCE.openai.azure.com/openai/deployments/YOUR-DEPLOYMENT/embeddings?api-version=2024-02-01
EMBEDDING_API_KEY=your-azure-api-key
EMBEDDING_MODEL=text-embedding-3-small

Mistral

EMBEDDING_PROVIDER=openai
EMBEDDING_URL=https://api.mistral.ai/v1/embeddings
EMBEDDING_API_KEY=your-mistral-api-key
EMBEDDING_MODEL=mistral-embed

Together AI

EMBEDDING_PROVIDER=openai
EMBEDDING_URL=https://api.together.xyz/v1/embeddings
EMBEDDING_API_KEY=your-together-api-key
EMBEDDING_MODEL=togethercomputer/m2-bert-80M-8k-retrieval

Local LLM servers (LM Studio, vLLM, etc.)

EMBEDDING_PROVIDER=openai
EMBEDDING_URL=http://localhost:1234/v1/embeddings
EMBEDDING_API_KEY=not-needed
EMBEDDING_MODEL=nomic-embed-text

Dimension Auto-Detection

mememory automatically detects the vector dimension of the configured embedding provider at startup. It sends a test string ("dimension probe") and measures the returned vector length.

On first run, the database is created with the detected dimension. On subsequent runs, the detected dimension is compared against the existing database column:

• Match — normal startup
• Mismatch — startup fails with an error explaining the conflict

This prevents silent data corruption from mixing embeddings of different sizes.

Switching Providers

Because different models produce vectors of different dimensions, you cannot simply change the provider and keep existing data. The vectors would be incompatible.

Migration process

1. Export your memories:

curl http://localhost:4200/api/memories/export -X POST -o backup.json

2. Stop the Docker stack:

docker compose -f docker/docker-compose.yml -p mememory down

3. Reset the database (deletes all data including vectors):

# Remove PostgreSQL data
rm -rf ~/.mememory/postgres

4. Update environment variables with the new provider configuration.

5. Start the stack:

docker compose -f docker/docker-compose.yml -p mememory up -d

6. Import the backup. This re-embeds all memories with the new provider:

curl http://localhost:4200/api/memories/import \
  -X POST \
  -H "Content-Type: application/json" \
  -d @backup.json

DANGER

The import re-embeds every memory through the new provider. For OpenAI, this costs tokens. For Ollama, this takes time proportional to the number of memories.

Implementing a Custom Provider

mememory uses a simple Embedder interface:

type Embedder interface {
    Embed(ctx context.Context, texts []string) ([][]float32, error)
    EmbedOne(ctx context.Context, text string) ([]float32, error)
}

To add a new provider, implement this interface and register it in the factory function. See the Contributing guide for details.