Use cases / Agent memory
Long-term memory that recalls the right fact for the right user.
Memories are full of names, dates and IDs that embeddings blur. WarpSearch stores each memory as one document with its text, vector and metadata, and recalls it with keyword and vector search together, filtered to one user and ranked by how recent and how important it is.
What a memory store needs
Hybrid recall
BM25 and vector search in one request, fused with RRF. "What is Dana's order number?" needs the exact term as much as the meaning. See Hybrid search.
Scoped to a user, agent or session
Keep user_id, agent_id and session_id as fields and filter on them. Filters apply inside the vector search, so one user's memories still return k results.
Recent and important first
A function_score query decays the score by age and lifts it by an importance field, so a fresh preference beats a stale one. See Function scores.
Searchable in the next turn
Write with commit: true and the memory is searchable when the call returns. See Consistency.
Facts that change
Store valid_from and valid_to dates and filter on them to recall what was true at a given time. Entities and relations can be graph edges, walked with Graph search (paid plans).
Read and write over MCP
Agents index and search memories through the MCP server with no client code. See the guide for AI agents.
What a memory backend costs
1M memories with 768-dim embeddings and 1M recalls a month, at list prices in US East.
| Platform | Small RAG |
|---|---|
| WarpSearch | $10 |
| turbopuffer | $161.6x cheaper |
| Weaviate | $48 to $1434.8 to 14x cheaper |
| Elastic Serverless | About $100 to $21010 to 21x cheaper |
| Pinecone | $118 to $13712 to 14x cheaper |
| OpenSearch Service | $132 to $39713 to 40x cheaper |
| OpenSearch Serverless | $357 to $53036 to 53x cheaper |
| Aiven | $365 to $73037 to 73x cheaper |
| Small RAG | 1M text chunks with 768-dim embeddings, 7 GB of JSON. 1M queries a month spread through the day. 1 GB written a month. |
|---|
List prices in US dollars, checked October 20, 2026. Ranges cover regions, plan choices, or how the vendor sizes compute. All comparisons and sources.
Good to know
WarpSearch stores and recalls memories; it does not decide what to remember. Fact extraction, summarization and contradiction handling stay in your agent or memory framework, and you compute embeddings yourself. Keep many users in one collection with a user_id filter rather than a collection each, since collections are capped per plan. Forget a memory or a user by deleting documents by id: there is no per-document expiry or delete by query yet.
Questions
What goes in a memory document?
The memory text, its embedding, the user, agent and session it belongs to, when it was created and last used, an importance score, and a type such as fact, preference or episode. Raw conversation turns and the facts extracted from them can live side by side, told apart by type.
How do I rank recent memories higher?
Wrap the query in a function_score with a gauss or exp decay on the created or last-used date, and a field_value_factor on importance. Keep the boosts small so relevance still decides between close candidates.
How do I forget a user?
Find the user's memory ids with a search filtered on user_id, then delete them by id in batches, with commit: true on the last one so they drop out of results right away.
Can I use it with Mem0 or another memory framework?
Memory frameworks do the extraction and keep the memories in a vector store, and WarpSearch speaks the Elasticsearch, OpenSearch and turbopuffer APIs they connect to. We have not yet tested a framework end to end, so contact us before you rely on one.
Read next
Free up to 1 GB, no card. Pay per GB and per request after that.