Letta AI

Letta AI builds software for AI agents with persistent memory. Founded by Charles Packer and Sarah Wooders, it grew out of UC Berkeley’s MemGPT research into an agent framework and hosted platform. In March 2026, Letta designated Letta Code its flagship and announced a shift toward an open, model-agnostic agent runtime combining computer use, git-backed memory, skills and subagents.[2][1][8]

That product direction followed earlier work on managing information outside a model’s limited context window. The research began before Letta’s September 2024 emergence from stealth.[4][3][8]

Packer and Wooders met during their PhD research at Berkeley’s Sky Lab under Joseph Gonzalez and Ion Stoica. Both professors joined Letta in advisory capacities. Letta traced MemGPT’s origin to an effort to improve Discord chatbots’ memory systems.[2][1]

MemGPT: Towards LLMs as Operating Systems, first submitted on October 12, 2023, introduced virtual context management. Inspired by operating-system memory hierarchies, it moves information between fast and slow memory tiers to provide extended context within an LLM’s limited context window, using interrupts to manage control flow between the system and user. The researchers evaluated it in document analysis and multi-session chat. The paper’s authors were Charles Packer, Sarah Wooders, Kevin Lin, Vivian Fang, Shishir G. Patil, Ion Stoica and Joseph E. Gonzalez.[4]

The project attracted attention before its planned code release. “Someone scooped us,” Packer told TechCrunch. He recalled that someone posted the October 12 whitepaper to Hacker News on Sunday, ahead of the founders’ intended Monday release of fuller research and code. TechCrunch’s September 2024 account reported that MemGPT had quickly attracted 11,000 GitHub stars and 1.2K forks—measures of project attention rather than customer adoption.[3]

By September 2024, the repository had expanded to custom tools, data sources, memory classes and agents deployed as persistent services. In its September 23 announcement, Letta resolved ambiguity around the MemGPT name: MemGPT would refer to the original agent design pattern with self-editing memory tools, while Letta would name the agent framework. The company committed to maintaining the open-source project, with commercial work focused on deployment, performance, debugging and monitoring.[1]

Letta announced $10 million in seed financing in September 2024, led by Felicis with participation from Sunflower Capital and Essence VC. TechCrunch reported a $70 million post-money valuation on September 23. The company’s September 24 release identified Packer as CEO and Wooders as CTO.[2][3]

On September 24, Letta released a free Agent Development Environment (ADE) and API platform and began onboarding early developers to the Letta Cloud beta. The hosted service was designed to run stateful agents through REST APIs while allowing developers to swap model providers. The ADE exposed agents’ prompts and memory for direct inspection and editing, including the exact context passed to the model at each reasoning step.[2]

Sleep-time Compute: Beyond Inference Scaling at Test-time, submitted April 17, 2025, explored preparing context before a query arrives: anticipating likely questions and precomputing useful information. Persistent context carries that preparation into later interactions. The research collaborators were Kevin Lin, Charlie Snell, Yu Wang, Charles Packer, Sarah Wooders, Ion Stoica and Joseph E. Gonzalez.[5][6]

The authors reported approximately 5× less test-time compute at matched accuracy on the modified benchmarks Stateful GSM-Symbolic and Stateful AIME. Scaling sleep-time compute increased accuracy by up to 13% and 18%, respectively. On Multi-Query GSM-Symbolic, sharing preparation across related queries decreased average cost per query by 2.5×. These were benchmark findings; the authors also found that effectiveness correlated with how predictable the user’s query was.[5]

Letta’s April 21 product explanation reported releasing sleep-time agents in Letta 0.7.0. The architecture separated a primary agent, which handled user messages, custom tools and external-memory searches, from a sleep-time agent, which edited in-context memory blocks. This allowed memory management to run asynchronously instead of sharing the conversational agent’s execution loop. The two agents could use different models and configurable execution frequencies.[6]

Letta’s December 16, 2025 announcement introduced Letta Code as a coding agent built around long-lived agents persisting across sessions. Its /init command prompted exploration of an existing project and formation of memory blocks; /remember explicitly triggered reflection. Reusable skills were stored as Markdown files that could be versioned in git or shared with other compatible agents. Letta reported that agents, with human engineers’ help, had contributed internal skills for database migrations, PostHog dashboards and API changes.[7]

On March 16, 2026, Letta made Letta Code its flagship and announced a consequential change in memory architecture. Specialized tools editing memory in a database would give way to general computer-use tools such as bash operating on git-backed context repositories, called “MemFS.” Sleep-time workflows would move from server-side agents to client-side subagents with access to the same computer, tools and context.[8]

The transition also called for replacing server-side filesystem features with actual filesystem access, templates with versioned SDK and community tooling, and server-side MCP integrations with client-side skills. Letta announced immediate deprecation of tool rules and some legacy tools, with larger features such as templates and filesystem scheduled for deprecation by mid-April. These were the announced replacements and timetable.[8]