Is Letta the same as MemGPT?

Letta is the platform that grew out of the MemGPT research project, carrying its core idea forward: the agent manages its own memory much as an operating system manages RAM and disk.

Can Letta agents use Brain?

Any agent that can call MCP tools can query a Brain with a scoped key, so an agent built on Letta can join a governed team brain alongside the rest of your fleet.

Which should a non-developer pick?

Brain. Letta is a platform for people building agents. Brain connects the tools you already use, with setup a non-developer can follow.

Compare

Brain vs Letta

An agent platform where the agent manages its own memory, and a governed brain your existing agents share.

Letta, from the MemGPT research lineage, is an agent platform: you build long-running agents that manage their own memory much as an operating system manages RAM and disk. Brain is not an agent platform. It is the governed memory the agents you already run connect to and share, with permissions and a verifiable record built in. Build a new agent on Letta. Give the agents you have a brain with Brain.

Join the waitlist Book a demo

Why teams compare them

Every reader, one grid
Self-managed memory decides what one agent keeps. It does not decide who else may recall it. Brain gives every person and every agent a named scope over the same knowledge.
Answers that show their working
When several agents share a memory, a recall without provenance is a claim nobody can check. Every answer carries the sources behind it.
Where Letta is the better fit
If you are building a long-lived autonomous agent as a product, you want the agent to own its memory decisions, and you value a self-hosted runtime with research behind it, that is Letta's home ground.

The questions people actually ask about Letta

Our agent manages its own memory. Isn't that better?

For one long-running agent built from scratch, quite possibly. Letting the agent decide what stays in context is the whole architectural bet, and it is a serious one.

It answers how one agent remembers. It does not answer who else may recall what it remembered, and for a team that second question arrives first.

Fine. But we could add access control around it ourselves.

You could, and on a self-hosted runtime you have every hook you would need.

It is a project rather than a setting, and it is the kind that is never quite finished: scopes, redaction, deletion you can prove, and a record that holds up when somebody asks for it. Brain ships that as the product, which is a different thing from shipping the parts.

Alright. But we would have to rebuild our agents on it.

On a runtime, yes, which is the honest cost of that architecture and worth paying when the agent is your product.

Brain makes the opposite assumption: the agents your team uses daily are already built by somebody else, and what they lack is a shared store of what your company knows. You point them at it; you do not rebuild them.

And the memories that shouldn't be shared with every agent?

Those turn up quickly once memory is extracted from real sessions: a figure, a customer, an unannounced plan.

Brain withholds the field rather than dropping the memory, so a scoped agent still gets the useful part, and the record says exactly what it did not get.

At a glance

Brain compared with Letta, attribute by attribute.
Attribute	Brain	Letta
What it is	A governed brain the agents you already run connect to over MCP	An agent platform with self-managed memory
Core idea	One shared, permission-aware memory for people and agents	The agent pages its own memory in and out, as an operating system manages RAM
What you bring	The agents you already use: Claude Code, Cursor, Codex, desktop assistants	Agents you build on the runtime
Access control	YesPer person and per agent, plus field-level redaction	NoAn application-level concern; the platform's focus is the agent's own memory
Audit and proof	YesTamper-evident, content-blind record, optional on-chain anchor, provable deletion	NoNot what the platform sets out to provide
Deployment	Hosted, per-tenant database isolation, bring your own model key	Open source and self-hostable, with a strong self-hosting story

Compared against publicly available information, last checked 20 September 2026. Products change. Tell us if something here is out of date: hello@heybrain.io

Who each one is for

Choose Brain if

Giving the agents you already run one shared, governed memory
Teams that need permissions, redaction and audit without building them
Mixed fleets: coding agents, desktop assistants and autonomous agents on one brain
Fast setup: minutes per agent, and no new runtime

Choose Letta if

Building long-running stateful agents from scratch
Research-grade control over what an agent keeps in context
Self-hosting the entire agent stack

We are not moving our agents onto a new runtime.

You would not be. Brain is not a runtime and has no interest in becoming one.

It is an MCP endpoint. Point an agent at a scoped key and it joins the team's memory; take the key away and it leaves. Nothing about how the agent is built has to change.

Join the waitlist

Questions, answered

Is Letta the same as MemGPT?: Letta is the platform that grew out of the MemGPT research project, carrying its core idea forward: the agent manages its own memory much as an operating system manages RAM and disk.
Can Letta agents use Brain?: Any agent that can call MCP tools can query a Brain with a scoped key, so an agent built on Letta can join a governed team brain alongside the rest of your fleet.
Which should a non-developer pick?: Brain. Letta is a platform for people building agents. Brain connects the tools you already use, with setup a non-developer can follow.