Letta vs Memlo: which to use for agent memory

Last updated October 31, 2026

Letta, formerly MemGPT, builds stateful agents: its open-source Letta Code harness runs agents whose memory, identity and skills live in a git repository that belongs to each agent, locally or in Letta Cloud. Memlo is a hosted memory server for the agents you already use: Claude Code, Codex, claude.ai, ChatGPT and other MCP apps share memories that agents save on purpose, each with its own version history. Pick Letta to build or run agents that carry their own memory. Pick Memlo to give the agents you use today, from several vendors, one shared memory.

At a glance

Facts about Letta come from its own site, repositories and documentation, read on October 31, 2026. Links are in Sources at the end.

How Letta works

Letta began as MemGPT, from the 2023 paper "MemGPT: Towards LLMs as Operating Systems," which proposed "virtual context management": moving information between a model's limited context window and slower storage, the way an operating system moves data between fast and slow memory. On September 23, 2024 the project announced "MemGPT Is Now Part of Letta."

Today Letta's main product is Letta Code, which its README calls "a stateful agent harness." Each agent's memory lives in MemFS: "The memory itself is part of the agent's state, held in a git repository that belongs to the agent." The agent reads and edits that repository with ordinary file tools, and "every memory edit is committed to the MemFS git repository," which "provides version history, conflict resolution, and a clear boundary between saved memory and uncommitted changes." When several cloud agents need the same knowledge, Letta has shared memory repositories owned by your organization.

Older Letta guides describe memory blocks, archival memory and the .af agent file. Letta's documentation now marks those as legacy, and the original Letta V1 API server is retired to an archive branch.

How Memlo works

Memlo is an MCP server at https://memlo.ai/mcp. You add it to each agent and sign in once. Agents get tools to save, search, read, change and retire memories.

• One memory for agents from different vendors. Claude Code, Codex, claude.ai and ChatGPT find the same memories when they sign in to the same account.
• History, with authors. Every change is a version, with when it was written and by whom: the person, and the connected app or agent they used. Any version can be restored, and edit_memory refuses a write when someone else changed the note first.
• Spaces and sharing. Keep memories in spaces, and share a space or a page with people by email as a Viewer or an Editor.

Where Letta is the better choice

• You want agents that carry their own memory and identity. That's what Letta is built for. Memlo adds memory to agents someone else built.
• You want memory as files in git. MemFS is a real repository the agent edits and commits, which suits developers who already think in git.
• You need to self-host. Letta runs locally or on your own infrastructure. Memlo is hosted only.
• You're building on an SDK. Letta has a TypeScript Agent SDK. Memlo has none.
• You want to start today. Letta has a free plan and open source code; Memlo is by invitation.

Where Memlo is the better choice

• You use agents from more than one vendor. Claude Code, Codex, ChatGPT and claude.ai read and write the same memories. Letta's memory belongs to Letta agents.
• You don't want to switch harnesses. Add one URL to the agents you already use, and sign in.
• You want to see which person and which app wrote each version. memory_history names both, and restore_memory rolls back.
• You share with people, not only agents. Share a space or a single page by email as a Viewer or an Editor.
• You want sign-in instead of keys. Every app signs in with OAuth and can be disconnected on its own.

What Memlo doesn't do yet

• Cursor. Memlo's sign-in doesn't work from Cursor yet.
• Agents. Memlo is memory only; it doesn't run agents, as Letta does.
• An SDK or API keys. Memlo is reached over MCP with your own sign-in.
• Self-hosting. Memlo is hosted only.
• Open sign-up. Accounts are by invitation for now; you can join the waitlist.

Frequently asked questions

Is Letta the same as MemGPT?

Yes. Letta's GitHub repository describes itself as "Letta (f.k.a. MemGPT)," and on September 23, 2024 Letta announced "MemGPT Is Now Part of Letta." MemGPT started as the 2023 research paper "MemGPT: Towards LLMs as Operating Systems."

Is Letta open source?

Yes. Letta Code is under the Apache 2.0 license, and Letta's documentation says you can "run Letta fully locally or on your own infrastructure — no Letta account required."

How does Letta store memory now?

In MemFS, a git repository per agent. Every memory edit is a commit. Memory blocks and archival memory are legacy features in Letta's current documentation.

How much does Letta cost?

On October 31, 2026, Letta listed Free ($0, up to 3 stateful agents), Pro ($20 a month, up to 20 agents), an API Plan ($20 a month, plus $0.10 per active agent a month and tool execution time), Teams Pro ($20 per seat a month) and Enterprise (custom).

Can Letta and Memlo work together?

Possibly: Letta agents can use MCP servers as tools, and Memlo is an MCP server with OAuth sign-in. We haven't tested a Letta agent against Memlo's sign-in.

Sources

Checked on October 31, 2026:

• Letta and Letta pricing
• Letta Code on GitHub and the letta repository
• MemFS and shared memory
• Letta hosted MCP server
• MemGPT: Towards LLMs as Operating Systems, arXiv, October 2023
• MemGPT is now part of Letta, Letta, September 23, 2024
• Memlo: the setup page, how it works and terms