Letta
Open-source framework for stateful AI agents with long-term memory
Last reviewed 2026-06-19
Letta is an open-source framework and platform for building stateful AI agents: agents that retain long-term memory and improve over time. It is the direct successor to MemGPT, the UC Berkeley research project whose 2023 paper introduced virtual context management, moving information between an LLM's limited context window and external memory. Letta's defining capability is self-editing memory: agents manage their own memory blocks (structured, labeled, individually persisted text that the agent can edit and share across agents), so they remember and evolve across sessions instead of starting stateless. The platform ships as a Letta agent server exposing a REST API (with Python and TypeScript SDKs), paired with the Agent Development Environment (ADE), a GUI to create, test, debug, and monitor agents with full visibility into the context window, memory state, and tool execution. Letta is model-agnostic and supports MCP tools, custom tools, and built-in tools. It is available open source under Apache-2.0 and as a hosted Letta Cloud, and is aimed at developers building agentic applications.
What it can do
Self-edit and manage long-term memory
AutonomousAgents decide when to read and write their own memory blocks across sessions, the core MemGPT-derived behavior, so they retain and evolve context.
sourceBuild and deploy stateful agents via API
CopilotExposes a REST API with Python and TypeScript SDKs so developers can create agents that act within their configuration.
sourceExecute tools including MCP and custom tools
SupervisedRuns built-in, custom, and MCP tools with human-in-the-loop support.
sourceInspect and debug agent state in the ADE
AssistantThe Agent Development Environment GUI gives full visibility into the context window, memory state, and tool execution for testing and debugging.
source
Strengths
- +Genuinely differentiated memory architecture (self-editing memory blocks from MemGPT research, not a RAG wrapper)
- +Open source under Apache-2.0, self-hostable, and model-agnostic, which avoids lock-in
- +Strong developer ergonomics: REST API, Python and TypeScript SDKs, and the ADE GUI with deep state visibility
Limitations
- −Positioning has become muddy across research lab, stateful-agents platform, and a Letta Code coding agent
- −Real usage costs climb well beyond the $20 Pro tier once LLM token pass-through is counted
- −Running stateful agents with persistent databases adds operational complexity versus stateless API calls
Overview
Letta is an open-source framework for building stateful agents that keep long-term memory. It is the successor to UC Berkeley's MemGPT project, which introduced moving information between an LLM's context window and external memory.
What it does
Its defining feature is self-editing memory: agents read and write their own memory blocks across sessions (the core autonomous behavior). Developers build and deploy agents via a REST API and SDKs (copilot-level developer tooling), run built-in, custom, and MCP tools with human-in-the-loop (supervised), and inspect agent state in the ADE GUI (assistant).
Integrations & setup
Self-host (Docker or pip) or use Letta Cloud. Model-agnostic across OpenAI, Anthropic, local, and Ollama; supports MCP and custom tools; ships Python and TypeScript SDKs and the ADE.
Pricing
Open source and free under Apache-2.0 to self-host. Letta Cloud is freemium with a Pro tier around $20/mo plus usage-based LLM pass-through, so real cost depends on token consumption.
Best for / not for
Best for developers who need agents with durable memory and deep state visibility. Less ideal for teams that want a fully managed, no-code product or who want to avoid running stateful infrastructure.
Alternatives
LangChain and LangGraph are the broad orchestration frameworks; CrewAI focuses on multi-agent workflows.
What people are saying
We aggregate real LinkedIn discussion into sentiment for the agents people search most. Letta isn't tracked yet, want it added? Request tracking.
FAQ
Is Letta the same as MemGPT?+
Letta is the company and platform that grew out of the MemGPT research project from UC Berkeley. MemGPT introduced the self-editing memory approach that Letta productizes.
What makes Letta different from other agent frameworks?+
Its focus is stateful memory: agents manage their own persistent memory blocks across sessions, rather than being stateless between calls. It is open source, model-agnostic, and ships an Agent Development Environment for debugging agent state.
Sources
- Letta (official site) · accessed 2026-06-19
- Letta documentation · accessed 2026-06-19
- Letta (GitHub) · accessed 2026-06-19
- MemGPT: Towards LLMs as Operating Systems (arXiv) · accessed 2026-06-19
Last reviewed 2026-06-19