Start free, then move to a monthly plan as memory API traffic grows. Writes and searches share one balance; reads stay free.
Teams that need a negotiated capacity or deployment review
Credits are consumed per API operation. The self-hosted open-source engine is free; infrastructure and model-provider costs remain with the deployer.
Inferred writes run fact extraction; infer=false stores submitted records verbatim. Reads and lists are free.
| Operation | Billing unit | Credits per request |
|---|---|---|
| memories.add (infer=true) | Started 8 KB input block | 2 |
| memories.add (infer=false) | Started 8 KB input block | 1 |
| documents.extract | Started 5 MB file block | 50 |
| documents.ingest | Projected source chunk | 1 |
| documents.search | Document search | 1 |
| memories.search | Semantic search | 1 |
| memories.get / list | Read | 0 |
A search costs 1 credit. Writes are measured in started 8 KB blocks of input: 2 credits per block for an inferred write, 1 per block for a verbatim one, so a normal conversational write is a single block. Reads and list operations are free. Document extraction is measured in started 5 MB blocks.
No. The Apache-2.0 engine is free to self-host. You remain responsible for your own infrastructure, model-provider, backup, and operations costs.
New writes and searches pause, while reads remain available. Upgrade the plan or contact support to resume metered operations.
Yes. Memory namespaces can be exported, and the related resources expose explicit deletion paths.
