Letta Review 2026

Framework and cloud service for building stateful AI agents that remember, learn continuously, and improve over time — the successor to MemGPT.

Letta is the commercial evolution of MemGPT, the influential UC Berkeley research project that showed how LLM agents could operate with layered memory (short-term working memory, long-term archival memory, and self-editing 'core memory') to escape the context window and behave more like a persistent assistant. Letta packages that idea as both an open-source Python framework and a hosted platform. Developers define an agent with a persona, a memory model, and a set of tools, then let it converse with users indefinitely — Letta manages what belongs in the prompt versus what belongs in archival memory, and the agent can literally rewrite its own core memory as it learns.

The framework is model-agnostic and works with Anthropic, OpenAI, Google, Groq, Together, and any OpenAI-compatible endpoint. It ships with a REST API and SDKs for Python and TypeScript, an Agent Development Environment (ADE) for visualizing memory and step-by-step reasoning, and integrations for building tools with function calling. Letta supports Model Context Protocol both as a client (agents can call MCP servers as tools) and as a server (Letta agents can be exposed to other MCP-aware apps). The hosted plan handles infrastructure and provides a UI for creating and managing agents; the open-source framework can be self-hosted. It is the pick when you specifically need long-lived, self-editing memory rather than a stateless chat.