Letta

What is Letta?

Letta is an open source framework and server for building stateful AI agents — agents that keep long-term, editable memory instead of starting fresh on every call. Formerly known as MemGPT, it treats the LLM like an operating system that manages its own context, so agents remember facts and conversations across sessions.

What is Letta best for?

Developers building agents that need to remember users, learn over time, and persist across many sessions — support assistants, personal AI companions, or research agents. It fits teams who want the agent runtime and memory to live on their own infrastructure rather than inside a hosted assistants API.

What can Letta do?

• Run agents as persistent services behind a REST API, each with its own saved state
• Manage layered memory: core memory the agent edits in real time, plus archival memory in a vector database for unlimited long-term recall
• Let the agent read and write its own memory using built-in memory-management tools
• Build and inspect agents visually in the Agent Development Environment (ADE)
• Work with any model provider — OpenAI, Anthropic, and local models via Ollama or vLLM
• Give agents tools and connect them to channels like Slack, Discord, and Telegram
• Call it from Python or TypeScript SDKs, or self-host the server with Docker

Where does Letta fall short?

• Agents run inside Letta’s runtime, so it’s a heavier commitment than a drop-in memory library — moving off it later means rewriting your agent infrastructure, not just a memory layer.
• It carries a real learning curve: you have to internalize its core/archival memory model and the agent-as-OS design before you’re productive.
• Memory quality depends on the model’s judgment. Because the agent decides what to save and retrieve, a weak model can drop information — and every memory operation spends extra inference tokens.

Is Letta free?

Yes — Letta is free and open source under the Apache 2.0 license, and self-hosting costs only your own server and model API usage. Letta Cloud is the optional managed service: a free tier of 5,000 monthly credits, a Pro plan at $20/month, and custom Team/Enterprise pricing for teams that don’t want to run the server themselves.

What does Letta replace?

Letta is a self-hosted alternative to hosted agent platforms like the OpenAI Assistants API, Microsoft Copilot Studio, and Amazon Bedrock Agents. It does the same stateful-agent job — persistent threads, memory, and tools — but the runtime and data stay on your infrastructure, and you aren’t tied to one vendor’s models.

FAQ

Is Letta open source? Yes. Letta is released under the Apache 2.0 license, one of the most permissive open source licenses, and the full framework and server are public on GitHub.

Is Letta the same as MemGPT? Yes — Letta is the successor to the MemGPT project from the same UC Berkeley team. MemGPT introduced the memory-management ideas; Letta is the production framework built around them.

Can I self-host Letta for free? Yes. Run the Letta server yourself with Docker at no license cost; you only pay for your server and whatever model API you point it at. Letta Cloud is the paid, managed alternative.

What do I need to run Letta? A server with Docker, a vector database for archival memory, and access to a model provider — a hosted API like OpenAI or Anthropic, or a local model served through Ollama or vLLM.