Letta, the platform behind MemGPT, is a framework and cloud service for building stateful AI agents with persistent memory that survives across conversations and sessions.
Most agent frameworks forget everything when a run ends. Letta treats memory as a first-class primitive: agents maintain editable, structured memory blocks that hold facts, preferences and task state, and the platform manages what is loaded into context at each step. The result is agents that remember users over weeks, not just within one thread.
Letta provides both an open-source framework and a hosted runtime with REST APIs, agent hosting, tool integration and observability. Because agents are exposed as stateful services rather than stateless prompts, you can build assistants that maintain long-running projects, learn from interactions and be inspected or edited like databases. Because memory lives outside the prompt, agents stay consistent even when the underlying model changes.
The open-source framework is free to self-host under a permissive license. Letta Cloud charges for hosted agent runtime usage, with a free development allowance and usage-based pricing that scales with model tokens and agent steps. Enterprise plans add dedicated infrastructure and support. Self-hosted deployments on your own infrastructure carry no platform fees at all.
Letta targets developers building assistants that must remember users, maintain long-running projects or run as persistent services. It is ideal for customer-facing agents, personal assistants and research agents where conversation history alone is not enough.
Yes, the open-source framework is completely free to run yourself, and Letta Cloud includes a free tier for development. Production workloads at scale move to usage-based or enterprise pricing.
Yes, the platform is API-first: agents are created, messaged and managed over REST endpoints, so they can be embedded in any application. The hosted API and the self-hosted server expose the same surface.
Compared with stateless orchestration libraries, Letta differentiates on memory: agents keep durable, inspectable state across sessions instead of re-reading documents each run. The trade-off is a learning curve around memory design versus simpler prompt-chaining frameworks. For a quick chatbot simpler stacks suffice, but for agents whose value comes from knowing the user, Letta’s memory model pays off immediately.