Using MetaMemory with Cohere Embed

Cohere has built its reputation on retrieval-first embedding models, and MetaMemory's integration takes full advantage of that specialization. The embed-v4.0 model represents the state of the art in embedding quality for search and retrieval tasks — it was explicitly trained to maximize the relevance of retrieved documents, which aligns perfectly with MetaMemory's memory retrieval pipeline. One of Cohere's standout features is native support for compressed embeddings. The embed-v4.0 model can produce binary and int8 quantized vectors that are dramatically smaller than full float32 representations while preserving most of the retrieval quality. MetaMemory leverages this to offer a compact storage mode that cuts memory costs significantly for high-volume deployments. The embed-multilingual-v3.0 model is another strong option, covering over 100 languages with a single model. Unlike some multilingual approaches that sacrifice English quality for breadth, Cohere's multilingual model maintains strong performance across all supported languages. MetaMemory automatically selects the right input type parameter (search_document for storage, search_query for retrieval) to ensure embeddings are optimized for their purpose. Cohere also provides explicit relevance scores alongside embeddings, which MetaMemory feeds into its adaptive strategy selection to further refine retrieval quality over time. For teams that prioritize retrieval precision above all else, Cohere is the provider to choose.

Setup Guide

Configuration Example

curl -X POST https://api.metamemory.tech/v1/providers \
  -H "Authorization: Bearer YOUR_METAMEMORY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "provider": "cohere",
    "api_key": "YOUR_COHERE_API_KEY",
    "default_model": "embed-v4.0",
    "settings": {
      "input_type": "search_document",
      "embedding_types": ["float"]
    }
  }'

Supported Models

Capabilities