Letta — the company that grew out of the MemGPT research project — builds stateful agents: you create agents inside the Letta server, and they manage their own memory as they run. It's a real framework with real ideas. But it's also a commitment: your agents live in Letta's runtime, built with Letta's abstractions.
The framework question
- Letta: agents run inside its server and get memory as a property of the framework. Great if you're starting fresh and want its whole model.
- Memniva: memory is a service your existing agents call — Claude Code, Cursor, a LangChain script, your own product — no runtime migration.
What Memniva adds around memory
- remember/recall/search as MCP tools and REST endpoints, with dedup and versioning.
- Document ingestion: PDFs, Office files, images (OCR), URLs with JS rendering — into the same retrieval system.
- Knowledge bases with hybrid RAG, cited chat, and live web search for fresh context.
Which should you pick?
If you want an opinionated agent runtime where memory management is the product, evaluate Letta seriously. If you want your current agents to remember — and to read documents, pages, and search results from one store — Memniva is the drop-in layer. Free to start.