memories embed

memories embed [options]

Generate vector embeddings for memories that don't have them yet. These embeddings power semantic search (memories search --semantic).

How it works

The command:

1. Finds all memories without embeddings
2. Downloads the embedding model if needed (~50MB for default model)
3. Generates vector embeddings for each memory
4. Stores embeddings in the local database

Everything runs locally — no API calls, no data leaves your machine.

Options

Examples

Generate embeddings for memories without them:

memories embed

Output:

Embedding 42 memories...

Model: all-MiniLM-L6-v2 (384d, fast)
First run downloads the model. Subsequent runs are faster.

✓ Embedded 42 memories

Preview what would be embedded:

memories embed --dry-run

Regenerate all embeddings (after switching models):

memories embed --all

When to run

Automatic

Memories added with memories add automatically get embeddings. You don't need to run memories embed for new memories.

Manual

Run memories embed when:

1. After importing memories — If you used memories import or memories ingest
2. After syncing from cloud — If memories were synced without embeddings
3. After switching models — Run memories embed --all to regenerate
4. First-time semantic search — If you have existing memories and want to use --semantic

Embedding models

The default model is all-MiniLM-L6-v2:

• Size: ~50MB download
• Dimensions: 384-dimensional vectors
• Speed: Fast (~10 seconds for 100 memories)
• Quality: Good for most use cases

You can switch to higher-quality models for better semantic accuracy:

# View available models
memories config model

# Switch to a better model
memories config model gte-base
memories embed --all

See Embedding Models for detailed comparison and recommendations.

Storage

Embedding storage depends on the model:

Troubleshooting

Model download fails

If the model download fails, check your internet connection and try again:

memories embed

The model is cached at ~/.cache/memories/models/, so partial downloads will resume.

Embeddings not working

If semantic search isn't returning results:

1. Check that embeddings exist: memories stats
2. Regenerate if needed: memories embed --all
3. Try a different model: memories config model gte-base

Large database

If your database is getting large, embeddings are the main contributor. Consider:

1. Using a smaller model (all-MiniLM-L6-v2 vs gte-large)
2. Using keyword search for simple queries
3. Only generating embeddings for important memories