Letta Review 2026

Stateful agent framework and cloud from the MemGPT team — agents that actually remember across sessions.

Letta (formerly MemGPT) is the open-source framework and hosted cloud for building agents with persistent, self-editing memory. Where most LLM apps forget everything the moment the context window rolls, Letta gives every agent a durable memory tier — core memory, archival memory, and message buffers — that the model reads and writes via tool calls. The result is agents that remember users across weeks of conversation, learn preferences, and can be paused, resumed, or migrated between models without losing state. Letta ships as a Python/TypeScript SDK, a self-hosted server, and a fully managed cloud at letta.com. The Agent Development Environment (ADE) is a browser IDE for inspecting an agent's memory, tools, and every reasoning step — invaluable for debugging long-running agents. Letta is model-agnostic (OpenAI, Anthropic, local Llama, Ollama), supports MCP for tool integration, and exposes agents through a REST API so they slot into existing backends. The team came out of UC Berkeley's Sky Lab with the original MemGPT paper and has become a de-facto standard for anyone building customer-support agents, personal assistants, or copilots that need to feel like they know you. Free self-hosted; Letta Cloud starts at $20/month with usage-based pricing for hosted agents.