Letta / MemGPT

Editorial note

Less a protocol than a memory architecture, but everyone in the category is converging on the rough shape Letta sketched.

Our analysis

Letta is not, strictly, an interop standard — it is a memory architecture and a runtime, originally derived from the MemGPT paper out of Berkeley. We include it here because the rough shape it sketches (tiered memory with explicit working-set, archive, and recall primitives) is the shape the category is converging on, and because Letta's API is one of the few that an interop standard for memory could be built against. The next twelve months will tell us whether the shape becomes a real standard (perhaps absorbed into MCP as a memory primitive) or whether it remains a runtime everyone clones the surface of. The interesting hire to watch is whether one of the larger model vendors commits to a Letta-shaped memory interface in their first-party agent runtime; that would force the question.

What we are watching

• Whether Anthropic or OpenAI's first-party memory APIs converge on Letta's shape
• MCP working-group conversation about a memory primitive
• Whether tiered memory becomes table stakes in agentic frameworks

How to cite this entry

Citation format: The Agentic Review. "Letta / MemGPT — Letta (formerly MemGPT) Memory Architecture." Standards Tracker, last updated 2026-04-30. https://agentic.review/standards/letta-memgpt/. See our citation style for additional guidance.

Update history

Material updates to this entry are dated and logged in our corrections log. The version above reflects the most recent review by the standards desk.