Letta

Letta is an open-source platform for building stateful AI agents, agents that remember across sessions, learn continuously, and improve themselves over time. It grew directly out of MemGPT, an influential research project from UC Berkeley that proposed treating a language model like an operating system: if the context window is limited, like a computer's RAM, then the model should be able to page information in and out and decide for itself what is worth keeping close at hand. Letta turns that idea into a general agent framework where persistence is the default rather than something bolted on afterward.

The heart of Letta is its tiered memory. Core memory is a small block that always lives in the context window, holding what the agent knows about the user and about itself, and the agent edits it directly using tools as it learns. Recall memory holds the searchable conversation history just outside the context window, and archival memory is long-term storage the agent queries when it needs older or larger knowledge. Crucially, the agent is an active manager of its own memory: it decides, through tool calls, what to keep in context, what to push to recall, and what to archive, rather than passively receiving whatever an external pipeline injects.

Letta is a full agent runtime, not just a memory layer. It manages the agent loop, tool execution, and state, persisting everything in a database so an agent runs as a long-lived service you can address rather than an object you rebuild on every request. It supports multi-agent setups where agents call sub-agents and share memory blocks, and it is model-agnostic across providers and local models, with Python and TypeScript SDKs and a full API.

More recently Letta has put much of its weight behind Letta Code, a memory-first coding agent that runs in the terminal and accumulates knowledge about a project across sessions, with a desktop app and connectors to chat platforms. Throughout, Letta's bet is the same: that durable, self-managed memory is what separates a long-running, genuinely useful agent from a stateless chatbot.