Letta
Letta is the platform for building stateful agents: AI with advanced memory that can learn and self-improve over time.
Last verified:
What is Letta?
Letta is an AI lab building persistent agents with the ability to continuously learn and adapt from their own experience. Born from MemGPT at UC Berkeley's Sky Computing Lab, Letta provides a memory-first platform for creating stateful agents that remember and learn over time. The core product, Letta Code, is a deeply personalized stateful coding agent that can learn your coding conventions and important codebase patterns over time.
Key features include persistent memory that allows agents to retain knowledge across sessions, an Agent Development Environment (ADE) for creating, testing, and monitoring agents with complete transparency into agent state, client-side tool execution for local Bash/Read/Write operations, support for multiple model providers (OpenAI, Anthropic, Google Gemini, zAI), skills system for extending agent abilities, subagents for multi-agent patterns, conversation threading for parallel sessions, and both desktop app and CLI interfaces. The platform also offers Constellation cloud gateway for accessing agents from any device.
Letta is designed for developers building coding agents, teams creating digital employees for tasks like writing reports and managing calendars, researchers working on AI memory and continual learning, and anyone wanting deeply personalized AI agents with unique identities. The platform supports building multi-agent systems with patterns like supervisor-worker, round-robin, parallel execution, and hierarchical teams.
Letta pricing
Pricing model: Freemium
Start using Letta Code for free by connecting your existing API keys and coding plans (like ChatGPT/Codex and zAI coding plans). You can create a free Letta account. The platform allows you to use your own API keys from external providers. Constellation cloud gateway is available for managing agent state and accessing agents from any device including chat.letta.com.
Letta pros
- Persistent memory across sessions
- Agent can learn from experience and improve
- Fully open source agent harness
- Supports multiple LLM providers (OpenAI, Anthropic, Gemini, zAI)
- Complete transparency into agent state via ADE
- Client-side tool execution for local environment
- Desktop app and CLI interfaces available
- Multi-threading with conversation support
- Skills system for extending agent abilities
- Subagents for complex multi-agent patterns
- Constellation cloud for device-agnostic access
- Background mode streaming for long-running tasks
- Resumable streams with crash recovery
- Direct memory editing and state control
- Shared memory blocks across conversations
- Python and TypeScript official SDKs
- Built-in retry logic and error handling in SDKs
- Can teleport agents across machines keeping memory intact
Letta cons
- Requires API key for API usage
- Need Node.js version 18+ for CLI installation
- Agent runs on external server (not fully local)
- Requires user approval before tool execution
- Desktop app currently only for macOS
- Learning curve for memory system customization
- Some features marked as experimental or legacy
- Deprecation of templates and versioning features
Frequently asked questions about Letta
What is Letta Code?
Letta Code is a deeply personalized stateful agent that can learn from experience and improve with use. It is a lightweight CLI harness built around the Letta TypeScript SDK that gives Letta agents the ability to interact with your local development environment. It attaches coding-focused tools and prompts to agents and provides a terminal interface for interacting with them.
How does the memory system work?
Letta uses a hierarchical memory system with core memory blocks (persistent knowledge the agent maintains), archival memory (external out-of-context memory store), and shared memory blocks across conversations. The agent can continuously learn and adapt from its own experience, with background memory agents that transform prompts, context, and skills over time.
What is the Agent Development Environment (ADE)?
The ADE is Letta's comprehensive toolkit for creating, testing, and monitoring stateful agents. It provides complete transparency into every aspect of agent operation including context window (memory, state, prompts) and tool execution. It has three main panels: Agent Simulator for testing, Agent Configuration for customization, and State Visualization for real-time insights into agent internal state.
What model providers are supported?
The Letta API supports a wide range of model providers including OpenAI, Anthropic, Google Gemini, zAI, and more. You can add your own LLM API provider keys via the /connect command in the CLI.
How does client-side tool execution work?
The core mechanism is that your Letta agent runs on an external server, but tools it calls like Bash, Read, and Write execute locally on your machine. When the agent requests a tool, the server pauses execution for approval, Letta Code executes the command locally, and the result is sent back to the agent to continue.
What are skills in Letta?
Skills are reusable modules that extend your agent's abilities. They can be learned through experience or pre-made. Letta released Skill Learning as a way to dynamically learn skills through experience, bringing continual learning to CLI agents.
Can I run multiple conversations simultaneously?
Yes, Letta Code organizes message threads into conversations. A single agent can have multiple conversations running in parallel, all sharing the same memory blocks and searchable message history. Use /new to start a fresh conversation, /resume to browse past conversations, or /conversation <id> to reopen a specific thread.
What is Constellation?
Constellation is a cloud gateway that manages your agent's state. Managing agents through Constellation allows you to easily access your agents from any device, including chat.letta.com. It enables remote control of agents running on any machine and teleporting agents across machines while keeping memory and context intact.
How do I get started with Letta Code?
Download the Letta Code desktop app for macOS or install via CLI (requires Node.js 18+). Open Letta Code and sign in with a free Letta account. Connect model providers via the bottom-left menu to add external API keys. The app starts with a default agent called Letta Code. Try asking your agent to explore your codebase or run /init to bootstrap its memory.
What multi-agent patterns does Letta support?
Letta supports multi-agent patterns including supervisor-worker, round-robin, parallel execution, producer-reviewer, and hierarchical teams. You can build these using the Letta API and SDKs for creating complex agent systems.