Overview
Letta (formerly MemGPT) is an open-source platform and managed cloud for building stateful AI agents — agents that keep persistent memory and improve during deployment, not just during training. Born from a UC Berkeley research paper, it gives an LLM an operating-system-like memory layer: a small core context, a searchable recall store, and a cold archival store, with the agent deciding what to promote or forget. Agents are exposed as REST APIs and SDKs (Python, TypeScript), run locally with no account, in the cloud, or on your own server. In 2026 the team shipped Letta Code, a memory-first coding agent ranked #1 among open model-agnostic agents on Terminal-Bench, and Channels for asynchronous agent messaging.
Key Features
- Three-tier memory (core / recall / archival) so agents retain facts and preferences indefinitely
- Model-agnostic: OpenAI, Anthropic, Mistral, or local models with state preserved across swaps
- Agent Development Environment to visualize memory, reasoning steps, and tool calls live
- REST API + Python/TypeScript SDKs; self-host (Apache 2.0) or managed Letta Cloud
- Letta Code desktop agent and Channels for async, multi-agent messaging
Pricing
| Plan | Price | For |
|---|---|---|
| Free | $0 | Up to 3 managed agents, bring-your-own-key, self-host |
| Pro | $20/mo | Up to 20 stateful agents plus Letta Auto model quota |
| Enterprise | Custom | Higher scale, SSO, dedicated support, compliance |
Comparison
Compared to LangChain, Letta centers on persistent memory and state rather than chaining tools, which matters once an agent must recall context weeks later. Compared to AutoGPT, Letta adds a production-grade memory store and a visual debugging environment instead of a bare autonomous loop.