The Berkeley spinout building stateful agents - AI that remembers everything, learns continuously, and improves itself over time.
Here is a fact about the AI chatbots most people use every day: they have no memory. You can spend an hour explaining your codebase, your preferences, your project - and the moment the conversation gets long enough, or you close the tab, all of it evaporates. The model does not remember you because, technically, it cannot. A large language model is stateless. Each request starts from nothing, reads whatever fits in a fixed window of text, and forgets the rest.
Letta, a San Francisco company spun out of UC Berkeley's Sky Computing Lab, is built on the wager that this is the actual bottleneck. Not that the models aren't smart enough - they are extraordinarily smart - but that they have no way to accumulate anything. Letta's founders, Charles Packer and Sarah Wooders, looked at the industry racing to build bigger and bigger brains and asked a quieter question: what if the frontier isn't a better brain, but a better memory?
This is a slightly unfashionable bet, which is part of what makes it interesting. The dominant story in AI is scale - more parameters, more compute, more data. Letta's story is about plumbing. It builds the system that sits around a model and manages what the model knows: which memories to keep in the context window, which to page out to a database, when to update them, and how to make all of that inspectable. If a chatbot is a person with amnesia, Letta is trying to give it a notebook, a filing cabinet, and the discipline to use both.
We are positioning ourselves as the open alternative to OpenAI.Charles Packer, Co-founder & CEO
The word Letta uses is stateful. A stateful agent, in their framing, is one where all of the state - memories, user messages, the agent's own reasoning, every tool call - is persisted in a database and never lost, even after it falls out of the model's short-term context. Important "core" memories get injected back into the window when needed. And crucially, the agent can edit its own memory through tools. It keeps a journal about you, and revises it.
The origin story is unusually clean. In October 2023, Packer and Wooders - PhD students who had met in Berkeley's Sky Lab under the same advisors, Ion Stoica and Joseph Gonzalez - posted a research paper called MemGPT. The idea was to treat a language model like a computer operating system: the context window is RAM, external storage is disk, and the agent pages memories between them the way an OS manages memory. It was a tidy, almost obvious framing, and the internet noticed. The paper went viral on Hacker News before it was even officially released. The open-source code has since grown past 23,000 GitHub stars.
Viral research papers do not always become companies, and companies do not always become good ones. But MemGPT had the useful property of describing a problem everyone building agents kept running into and nobody had cleanly solved. In September 2024, Letta emerged from stealth with a $10 million seed round led by Felicis at a $70 million post-money valuation. The cap table reads like a who's-who of people who know something about AI infrastructure: Jeff Dean, the chief scientist at Google DeepMind; Clem Delangue, CEO of Hugging Face; Cristobal Valenzuela of Runway; and the CEOs of MotherDuck, dbt Labs, and Hex, among others.
Led the MemGPT research toward an "operating system for LLMs." Frames Letta as the open counterweight to closed agent platforms. Berkeley PhD, advised by Ion Stoica and Joey Gonzalez.
PhD in Computer Science from UC Berkeley, focused on systems for AI. Builds agents that learn over time through memory that is model-agnostic and interpretable.
The trick isn't magic - it's bookkeeping done well. Letta learns by editing the context window, not the model's weights. That keeps the learning readable and portable across any model.
You talk to the agent. Messages, reasoning, and tool calls all flow through the context window.
→Every piece of state is written to a database, so it survives even when evicted from the window.
→The agent uses tools to update its own "core" memories - deciding what about you is worth keeping.
→Relevant memories are paged back into context when needed. The agent picks up where it left off.
Sleep-time compute: thinking while idle
In April 2025, Letta published one of its more counterintuitive ideas: agents should do heavy reasoning during their downtime, not at the moment you ask a question. They call it sleep-time compute. A dual-agent setup runs one agent for live conversation and a second "sleep" agent that activates during idle periods to reorganize memory, parse documents, and pre-process context. On benchmarks like GSM-Symbolic and the AIME math exam, shifting work to downtime cut test-time workload by up to five times without hurting accuracy.
Git-based memory
In February 2026, Letta rebuilt agent memory around what it calls context repositories: git-based versioning for what an agent knows. You can branch it, diff it, and roll it back. When an agent learns something wrong, you don't retrain a model - you check out an earlier version of its mind.
Letta is open-core: a free framework for adoption, a hosted cloud and enterprise offering for scale. Here is the toolkit.
The platform for stateful agents. Memories, messages, reasoning, and tool calls are persisted in a database so nothing is lost when it leaves the context window.
Build, deploy, and scale stateful agents with advanced memory systems - without managing the underlying infrastructure yourself.
A full-featured REST API with Python and TypeScript SDKs to drop memory-enabled agents into your own applications.
An open coding agent harness that learns from experience and gets more useful the longer you work with it. Install: npm i -g @letta-ai/letta-code
Named case-study customers include:
Use cases range from personalized chatbots and customer-support bots to healthcare symptom-tracking - anywhere an agent needs to remember across sessions.
Announced September 2024, led by Felicis (Astasia Myers), with Sunflower Capital and Essence VC.
Valuation at the seed - notable for a company only months out of stealth.
GitHub stars on the open-source repo, plus 2,500+ forks. Community is the top of the funnel.
The paper introducing self-editing memory for LLMs spreads on Hacker News before its official release.
Letta announces a $10M seed led by Felicis at a $70M valuation and unveils plans for Letta Cloud.
A new scaling direction: agents reason during idle time, cutting test-time work up to 5x on math benchmarks.
Research on learning by updating context rather than model weights - interpretable and model-agnostic.
Context repositories rebuild agent memory around programmatic management and versioning.
MemGPT reframes a language model as a computer: the context window is RAM, external storage is disk, and the agent pages memories in and out.
Letta's agents can edit their own memory through tools - essentially keeping and revising a private journal about you.
The founders met as PhD students under the same two Berkeley advisors, Ion Stoica and Joey Gonzalez.
Brand slogans include "Viva la machina" and "Context is selfhood" - a research lab that isn't afraid of a slogan.
The logo is a square nested inside a square: a self-contained memory core that holds its own state.
Sleep-time compute was developed in collaboration with Ion Stoica, co-founder of both Databricks and Anyscale.
Demos, docs, and the open-source code - Letta does most of its thinking in public.
letta.com · letta.com/about-us · github.com/letta-ai/letta · docs.letta.com · TechCrunch (Sep 2024) · PR Newswire · Crunchbase · letta.com/blog/sleep-time-compute · Fast Company. Funding, valuation and figures reflect publicly reported figures as of the seed round (Sep 2024) and Letta's published research; some metrics (GitHub stars, employee count) are approximate.