AI Agent Memory / RAG Cost Calculator v2
Estimate embedding, vector storage, query, reranker, long-term memory, and change-data-capture costs for agent memory and RAG.
Corpus & workload
Cached queries skip the vector DB search and reranker.
Cheap default; 1536 dims.
Pay for stored vectors plus read units.
Per-query reranking of top-k results.
Total 12-month cost
—
RAG + CDC + long-term memory
One-time embedding
—
Monthly total
—
Storage / mo
—
Query cost / mo
—
Reranker / mo
—
CDC / mo
—
Long-term memory / mo
—
Cost per 1k queries
—
Effective chunks
—
Vector storage
—
Prices are directional estimates. Check each provider's current pricing before budgeting.
Vector DB cost comparison
Same corpus, embedding model, query volume, and CDC cadence across all databases. Includes storage, uncached query search, query embedding, reranker, and CDC re-embedding costs.
| Vector DB | Storage/mo | Search/mo | Query embed/mo | Reranker/mo | CDC/mo | Total/mo |
|---|---|---|---|---|---|---|
| Pinecone Serverless managed | — | — | — | — | — | — |
| Weaviate Cloud managed | — | — | — | — | — | — |
| Qdrant Cloud managed | — | — | — | — | — | — |
| Zilliz / Milvus Cloud managed | — | — | — | — | — | — |
| Redis Cloud (vector search) managed | — | — | — | — | — | — |
| pgvector (RDS / self-hosted) self-hosted | — | — | — | — | — | — |
| Chroma (self-hosted) self-hosted | — | — | — | — | — | — |
How the estimate works
- Chunks: total corpus tokens ÷ effective chunk size after overlap.
- Storage: each chunk stores a vector (dims × 4 bytes) plus metadata.
- Embedding cost: one-time charge to embed all corpus tokens, plus per-query embedding tokens for uncached searches.
- CDC: recurring cost to re-embed a percentage of the corpus at the chosen cadence.
- Long-term memory: per-turn cost applied to memory turns stored each month.
- Cache: cached queries skip search and reranker entirely.
Where to cut costs
- • Use a cheaper embedding model if retrieval quality allows it.
- • Raise cache hit rate with query embedding cache or response cache.
- • Self-host pgvector or Chroma if you already run a server.
- • Batch CDC updates weekly instead of daily to reduce re-embedding.
- • Skip rerankers for internal prototypes; add them after product-market fit.
- • Reduce dimensions or quantize vectors if your DB supports it.