MIND vs Letta

An open-source framework, formerly MemGPT, for building stateful AI agents with self-editing memory — agents manage their own memory blocks directly, persisted across sessions, with a visual Agent Development Environment (ADE).

An open-source framework, formerly MemGPT, for building stateful AI agents with self-editing memory — agents manage their own memory blocks directly, persisted across sessions, with a visual Agent Development Environment (ADE).

Letta is the direct descendant of the MemGPT research project — the paper that first proposed treating LLM context like an operating system treats RAM: a small, fast core context the model always sees, and larger recall and archival storage the agent pages information in and out of via function calls, the way a real OS pages memory to disk. That framing is genuinely influential; a lot of what agent memory means in 2026 traces back to it. Letta is the maintained, production version of that idea — Apache 2.0, free to self-host, with an actual visual tool, the ADE, for watching an agent's memory blocks change in real time.

The core mechanic, self-editing memory, where the agent calls a function to update its own memory block, is powerful and also puts the work on the agent, and the developer building it, to decide what's worth remembering and how to structure it. MIND makes that call on the memory's behalf, using the same kind of pattern-noticing a person's own memory does, rather than requiring an agent to explicitly decide it should update its memory block now. Letta is a runtime you build an agent inside of; MIND is memory you point an agent, or a person, at, without writing memory-management logic yourself.

Where the two genuinely rhyme: both reject the idea that memory should just be shoved into a vector database and found later by similarity search alone. Letta's answer is structured, developer-managed memory blocks with an OS metaphor; MIND's answer is a graph that gets built for you and is reachable by 50-plus models, not just the agent framework you built it in.

Letta's memory blocks are structured text you can export via its API; that text imports into MIND as documents cleanly, though the block-editing mechanics themselves, the functions an agent calls to update its own memory, don't have a MIND equivalent, since MIND updates its graph automatically rather than through agent-invoked edits. Most people running agents on Letta keep it for the agent's own working memory and use MIND for the memory a person directs.