Letta is a framework for stateful agents and the successor to the MemGPT project. The core idea is that the agent does not restart with every conversation: it keeps long-term memory and decides what to store, revise or forget.
What it is for
- Building assistants that remember user preferences, decisions and history
- Keeping context across long tasks that span several sessions
- Separating working memory, long-term memory and external knowledge
- Packaging a complete agent, with memory and tools, to run elsewhere
How it works
- The agent is defined with editable memory blocks, tools and data sources
- The agent itself rewrites its memory blocks as it learns
- State is persisted in a database with a versioned message history
- It ships a TypeScript SDK and its own server, with MCP support for tools
- An agent file lets you export and import the agent between environments
Availability
- Open-source core under the Apache-2.0 licence
- The company also offers a desktop app and a managed cloud service
- The old V1 server is retired; new projects use the current line
Points to note
- Persistent memory requires a well-defined retention and privacy policy
- Memory-heavy agents can get expensive in tokens; watch consumption
- It is a code framework aimed at developers