Letta vs LangGraph
Letta (formerly MemGPT) is a stateful-agent platform: each agent keeps long-term memory in a git-backed file store and delegates to subagents, driven through a TypeScript Agent SDK, CLI or App Server. LangGraph is a Python graph runtime where you design state, memory and control flow yourself. Letta is a ready harness; LangGraph is a building kit.
Facts side by side
| Fact | Letta | LangGraph |
|---|---|---|
| Type | Framework | Framework |
| Languages / SDKs | TypeScript, Python | Python |
| License | Apache-2.0 | MIT |
| Pricing model | Open core | Open core |
| Orchestration pattern | Supervisor | Graph |
| GitHub stars | 24,981 (as of 2026-09-30) | 42,511 (as of 2026-09-30) |
| GitHub forks | 2,637 | 7,201 |
| Last push | 2026-09-10 | 2026-09-29 |
| Latest release | 0.16.8 | cli==0.4.32.dev0 |
| Repository | letta-ai/letta | langchain-ai/langgraph |
| Website | www.letta.com | www.langchain.com |
| Documentation | docs.letta.com | docs.langchain.com |
| Last verified | 2026-09-30 | 2026-09-30 |
| MCP support | Yes (checked 2026-09-30) | Yes (checked 2026-09-30) |
| A2A support | Unknown (checked 2026-09-30) | Yes (checked 2026-09-30) |
| AG-UI support | Unknown (checked 2026-09-30) | Partial (checked 2026-09-30) |
| Install verified | Not yet | 2026-09-30 |
Choose Letta if
- You want persistent agent memory managed for you, versioned as Markdown files in a git repository (MemFS).
- You build in TypeScript and want an SDK that creates an agent once and resumes sessions later.
- You want a ready agent harness with subagents and permission modes rather than a graph to design.
Choose LangGraph if
- You are a Python team: Letta's Agent SDK is JavaScript/TypeScript only.
- You need custom control flow with conditional edges, parallel branches and step-level checkpoints.
- You need an A2A endpoint; no official A2A support was found for Letta.
Migration notes
These tools sit at different layers, so migration is a redesign. Note first that the letta-ai/letta repository is now a landing page: the V1 Python server is retired and receives no fixes; current code lives in Letta Code and the Agent SDK. Guides that use the Python server packages or letta/letta Docker images target the retired server. Going from LangGraph to Letta, graph state and store contents become MemFS files, and nodes that called sub-agents become Letta subagents (each runs in its own subprocess and returns only its final message). Going from Letta to LangGraph, you must design a state schema, choose a checkpointer and a store for long-term memory, and rebuild tool permission prompts with interrupt().
FAQ
Is Letta still the MemGPT Python server?
No. The repository's SECURITY.md states the V1 server is retired. The current product is the Letta agent harness (Letta Code) with a TypeScript Agent SDK.
Can Letta be self-hosted?
Yes. The docs describe a local runtime and a self-hosted App Server that need no Letta account. Some features, such as dynamic workflows, are documented for Letta Cloud only.
Which one has built-in long-term memory?
Letta: memory is a core feature (MemFS). LangGraph provides stores for cross-thread memory, but what to save and when is your code.