MIND vs MemGPT

The research project and open-source framework (2023) that first proposed OS-inspired virtual context management for LLM agents — paging facts between a small always-visible context and larger recall/archival storage. The codebase is now maintained as Letta.

The research project and open-source framework (2023) that first proposed OS-inspired virtual context management for LLM agents — paging facts between a small always-visible context and larger recall/archival storage. The codebase is now maintained as Letta.

If you're reading this page, you're probably searching MemGPT out of habit — it's still the name a lot of the agent-memory conversation uses, even though the project itself has been renamed and folded into Letta. The idea MemGPT introduced in 2023 is worth understanding on its own terms, separate from whatever the current product is called: LLMs have a small, fixed context window, so treat that window the way an operating system treats RAM — keep the most relevant things in core context, and let the agent explicitly page other facts in from recall and archival storage via function calls when it needs them. That's a genuinely elegant answer to a real constraint, and a huge share of what agent memory means today traces back to that paper.

What's changed since is mostly branding and maturity, not the core idea: the original repository is now effectively a landing page pointing to Letta, where the OS-inspired memory tiers live on as self-editing memory blocks with a real product, an ADE for visual debugging, and Apache 2.0 self-hosting. MIND didn't descend from this lineage at all — it wasn't built to solve an LLM's context-window constraint, it was built to be a person's memory that any model can read. Those are related problems, both are versions of how do you not lose what mattered when context is limited or resets, but MIND's answer doesn't involve an agent explicitly deciding to page a fact in or out; it decides what mattered when you add it, the way your own memory works, not the way a paging algorithm works.

So the honest version of MIND versus MemGPT is really two comparisons in one: against the original research idea, MIND is a different design philosophy, automatic and person-facing rather than agent-managed and developer-facing. Against what MemGPT actually is today, the fair comparison to run is MIND versus Letta, which is its own page on this site with current pricing and features, because that's where the maintained code and the real product both live now.

There's no migration path from the archived MemGPT repository itself; anyone still running the original code should look at Letta's own migration guidance first, since that's where active maintenance lives. From either, memory content exported as text imports into MIND as documents the same way any other memory-tool export does.