Letta

Letta, formerly MemGPT, is an open-source framework and cloud platform for building stateful AI agents with persistent, operating-system-style memory. It manages tiered memory (core, recall, archival) so agents retain and evolve knowledge across sessions. The framework is free to self-host, while Letta Cloud offers a managed backend with free and paid tiers (Pro around $20/month as of 2026).

Letta, formerly known as MemGPT, grew out of a 2023 UC Berkeley research paper arguing that LLM agents should not rely solely on the context window for memory but instead use an operating-system-style memory hierarchy. Letta implements this with tiers: Core Memory that lives in context (like RAM), Recall Memory of searchable conversation history (like a disk cache), and Archival Memory for long-term storage the agent queries via tools (like cold storage). The result is agents that persist and evolve their memory across sessions rather than starting fresh each time. Letta is offered as an open-source framework you can self-host and as Letta Cloud, a managed platform for creating and running stateful agents with a hosted memory backend, an agent development environment, and APIs. This gives the standard open-source-versus-managed split: run the framework yourself for free, or use Letta Cloud's free and paid tiers for a hosted experience. As of 2026, Letta Cloud advertised a free tier (a small number of managed agents, bring-your-own-key), a Pro plan around $20/month, and custom team/enterprise pricing. Letta is a focused, credible option specifically for the memory problem in agents, backed by the framework's research pedigree. It is narrower than general agent frameworks, so it is often used alongside or within them, and, like others in the fast-moving agent-memory space, its cloud pricing and features continue to evolve.