Letta

Market position

Letta positions itself as an AI research lab whose research ships as software, and as an open, model-agnostic alternative to closed provider stacks — its CEO described the company as "positioning ourselves as the open alternative to OpenAI." Its GitHub presence (24.3k stars, 2.6k forks) and origin as the MemGPT project give it visibility in the agent-memory infrastructure category. TechCrunch noted competition from LangChain and from model providers' own agent tooling such as OpenAI's Assistants API, and raised OpenAI's o1 multistep model as a potential reduction in the need for external state management.

Differentiators cited in the sources are model agnosticism (agents portable across OpenAI, Anthropic, Mistral and self-hosted models, versus provider-locked tooling), open source licensing and transparency of memory state ("white-box memory") in contrast to frameworks that obscure prompts and memory, persistence of agent state and identity across sessions and devices, and skill learning that converts past experience into reusable, git-versionable artifacts.

Technology

Core technology is memory management for LLM agents: virtual context management (MemGPT), self-editing memory in which the model updates its own memory blocks, "white-box memory" giving developers visibility into and control over the exact prompts and memories passed to the model at each reasoning step, and sleep-time compute in which agents reason about context during idle time rather than at inference. Later work describes programmatic context management with git-based versioning of context repositories, continual learning in token space, and memory models trained with memory-native reinforcement learning. The platform includes an agent harness, terminal UI, App Server runtime, TypeScript Agent SDK, skills stored as markdown files, schedules, mods, and vector, full-text and hybrid search over messages and tools. The stack is model-agnostic, letting developers swap LLM providers and even run a single agent across multiple models.

Go-to-market

Distribution is developer-led and open source-first: the harness is installable from npm (npm install -g @letta-ai/letta-code), the source is on GitHub under Apache-2.0, and documentation, a Discord community and X presence support adoption. Paid/managed use is channeled to Letta Cloud, with desktop and browser apps and Slack/Telegram/Discord channel integrations extending reach to non-CLI users. The company also publishes research posts and case studies (Bilt, 11x, Kognitos, Hunt Club) and has received press coverage in WIRED, Fast Company and TechCrunch.

Developers and engineering teams building coding agents, personal assistants, AI coworkers and stateful agents embedded in applications; enterprise users are indicated through featured case studies with Bilt, 11x, Kognitos and Hunt Club. Earlier coverage cited personalized chatbots and long-running assistants such as symptom-tracking health bots as use cases.

Geography

Headquartered in San Francisco; sources describe no other offices.

History

Letta originated in UC Berkeley's Sky Computing Lab, the successor to RISELab and AMPLab and the source of projects including Spark, Ray, vLLM, Gorilla LLM and SGLang. The MemGPT whitepaper was posted on October 12, 2023 and was shared on Hacker News before the planned release, staying at the top of the site for 48 hours and driving rapid open source adoption. The company emerged from stealth on September 23-24, 2024 with a $10M seed round and released its Agent Development Environment and API platform, opening a beta waitlist for Letta Cloud. Subsequent research on sleep-time compute (2025) and continual learning preceded the December 2025 launch of Letta Code. The original letta-ai/letta repository was later converted into a landing page, with the retired V1 server source preserved on an archive branch and active development moved to letta-ai/letta-code.

Risks & controversies

Reported competitive risks include established players such as LangChain and agent tooling from major model providers, and the possibility that multistep reasoning models that maintain state internally reduce demand for external memory layers. The retired Letta V1 server source on the archive branch is explicitly unsupported, receives no fixes or security updates, and is stated to be unsuitable for production. Research posts dated into 2026 on the company site cannot be independently corroborated by the other sources provided.