💻
Freemiumai agentsmemoryopen sourcelong-term memoryself-improving

Letta (formerly MemGPT) is a framework for building AI agents that manage their own memory — deciding what to store, recall, and forget across conversations and sessions. Agents built on Letta can improve over time as their memory cores are updated by experience, making them suitable for long-horizon tasks and personalized assistants. The framework is open source and supports deployment via a REST API, with Letta Cloud offering a managed hosted option. It is popular in research and enterprise AI agent projects.

What is Letta?

Letta, formerly MemGPT, is a framework for building agents that manage their own memory. The agent decides what to store, what to recall, and what to let go across conversations and sessions, so its memory core is updated by experience rather than reset every time a context window fills. That makes it aimed at long horizon work and personalized assistants, where the useful state of the system is everything it learned last week rather than the last few messages. It is open source and agents are deployed behind a REST API, which means the agent becomes a service your application calls instead of a loop living inside one script. Letta Cloud is offered as a managed hosted option for teams that do not want to run it themselves. The catalog notes it is used in research and enterprise agent projects.

Who is Letta for?

Letta is for developers building a product where the agent has to remember. A support assistant that should know a customer's history, a research companion that accumulates findings over weeks, or an internal tool whose value grows with the context it has collected are the cases that justify the extra architecture. It fits when you have decided memory is a first class concern rather than something you will approximate by stuffing more text into a prompt. It is too much machinery for a one off script, for stateless generation tasks, or for anyone who wants a finished assistant to use rather than a framework to build with. Expect to write code and to own the deployment.

How does Letta compare?

CrewAI is about coordination rather than recall. It helps you define multiple agents with roles and handoffs, which solves a different problem, and the two are often complementary rather than competing. AgentOps sits alongside both as the observability layer, tracing what your agents did, what they cost, and where they failed, which is what you reach for once something is running and misbehaving. Hermes Agent is the closest in spirit on the memory question, since it also learns from interaction history, but it ships as a usable local agent rather than a framework you build a product on. Choose Letta when persistent, self managed memory is the feature your application is actually selling.

Compare with: CrewAI, AgentOps, Hermes Agent

Letta pricing

The catalog records Letta as freemium. The framework itself is open source, so you can self host and pay nothing beyond your own infrastructure and whatever model calls your agents make. Letta Cloud is the managed path for teams that would rather not operate it, and that is where paid plans come in. This is a healthy arrangement for evaluation, because you can prototype locally at no cost and only consider hosting once the agent is doing something worth keeping online.

Notes

No notes have been added for this tool yet.

Letta FAQ

Is Letta the same as MemGPT?
Letta is the current name of the project formerly called MemGPT. The core idea carried over, which is agents that manage their own memory by deciding what to store, recall, and forget instead of relying entirely on whatever fits in a single context window at request time.
Can I self host Letta?
Yes. The framework is open source and agents deploy behind a REST API, so you can run the whole thing on your own infrastructure. Letta Cloud exists as a managed option for teams that prefer not to operate it, which is the paid side of the freemium listing in this catalog.
When is memory worth the extra complexity?
When the product's value depends on what the agent learned earlier. Support assistants, research companions, and long running internal tools qualify. For stateless generation, a single prompt with the right context attached is simpler and cheaper, and adding a memory framework mostly adds moving parts you have to maintain.
Browse all Code & Dev