If you're building an AI product, your users expect it to remember them — and building memory infrastructure yourself means a vector store, an embedder, dedup logic, and an API on top. A memory API gives you the finished layer: three endpoints your backend calls, with retrieval quality that's someone else's full-time job.
The core operations
- remember — store a fact with provenance; near-duplicates are detected and merged instead of piling up.
- recall — retrieve by meaning: send the user's question, get the relevant memories ranked.
- search — hybrid semantic + keyword search across memories and ingested documents alike.
What surrounds the endpoints matters
- A typed SDK (@kitforai/sdk) so integration is an import, not a REST client you maintain.
- MCP parity: the same memory your API writes is available as native tools in Claude, Cursor, and any MCP agent.
- Document ingestion in the same account — memory records and reference documents share one retrieval system.
When to use an API vs MCP
Use the API when your code owns the loop — a SaaS backend, a pipeline, a custom agent. Use the MCP server when an existing agent should manage memory itself. They read and write the same store, so you can do both.
Ship memory this week instead of this quarter — Start free and make your first remember call.