Letta Review, Pricing & Alternatives (2026)

Letta is an agent platform built around persistent state and memory, developed by the team behind the MemGPT research project at UC Berkeley. The company launched publicly in September 2024 and raised a $10 million seed round led by Felicis, according to TechCrunch and Felicis. Its open source agent framework is available under the Apache 2.0 license. Documentation covers coding agents, personal assistants, AI coworkers, and agents placed inside applications. Users can work through the web app, CLI, Agent SDK, hosted Letta Cloud, or a self managed Letta App Server. The central product choice is persistent, inspectable agent state. Instead of treating every exchange as a fresh prompt, Letta keeps memory, tools, instructions, and conversation context available for later work.

Letta divides memory into editable in context blocks, recall memory for conversation history, and archival memory backed by searchable storage. A developer creates an agent, chooses a model, adds instructions and tools, then sends messages through the web app, CLI, SDK, or REST API. The agent can call tools, search stored information, and update its state. The Agent Development Environment shows system instructions, tools, memories, messages, reasoning, and context allocation for inspection and editing. Developers can set limits for memory sections and tool responses. Letta uses recursive summarization and message pruning to control context size. Blocks can be queried, edited, and attached to other agents, giving developers a direct way to share selected memory while retaining control over what remains in context.

Letta is model agnostic, so developers can change model configurations without replacing memory or agent state. In April 2025, Letta released sleep time agents, pairing a primary agent with a second agent that reorganizes memory during idle periods. In September 2025, its Memory Omni Tool allowed agents to create, delete, and reorganize memory blocks while learning. Claude Sonnet 4.5 was specifically trained for that tool, though the tool also works with other models available through Letta. Documentation presents the CLI as a coding and personal agent framework with local files, integrations, recurring schedules, and skills. The service therefore joins agent orchestration with a memory layer that stays accessible through APIs, while the visual development environment provides a place to inspect reasoning, tool calls, and context decisions during development.

Letta’s Free plan costs $0 per month and includes up to three stateful agents, 5 MiB of storage per memory repository, limited Letta Auto usage, and bring your own key support. Pro costs $20 per month, includes up to 20 stateful agents, 50 MiB per repository, Letta Auto quota, remote sandboxes, and paid overage. The API Plan costs $20 per month, supports unlimited agents, and charges $0.10 per active agent monthly, $0.00015 per second for server side tool execution, plus model costs. Teams Pro costs $20 per seat monthly. Enterprise pricing is custom and adds higher quotas, role based access control, SAML or OIDC single sign on, and dedicated support. Client side tools run on the user’s machine without Letta execution charges, but model use and heavy tool workloads can raise total spending.

Letta’s memory centered design asks developers to inspect, test, and govern state as agents change it over time. Hosted services collect data used to provide those services, including message requests and responses, and the privacy policy says such data may improve services or train future models. When a user runs a model backend independently, Letta says it does not collect those messages. LangGraph is a closer fit for developers who want explicit state graphs and workflow control, while Mem0 concentrates more narrowly on memory infrastructure. Mastra offers a broader TypeScript framework for agents and applications, and AutoGen and CrewAI focus on multi agent coordination. Choose Letta when persistent, editable agent memory is central, especially if you are comfortable managing context and state; choose another tool when workflow graphs or broad application scaffolding matter more.