Letta (MemGPT)
Open source runtime for stateful agents: instead of an agent restarting from scratch on every call, Letta manages its memory the way an operating system manages RAM and disk. A direct descendant of the MemGPT research paper, it tiers memory three ways: working memory in context (the RAM), searchable history outside context (the disk cache), archives queried on demand (cold storage). The agent itself decides what to keep at hand and what to archive, via function calls. You code the agent in Python or TypeScript, or drive it from the Letta Code CLI.
Strengths
- Three-tier memory modeled on an OS's virtual memory (context, history, archives)
- The agent manages its own memory: it decides what to keep, what to archive
- Full runtime (CLI + Python and TypeScript SDK), not just a layer to plug in
Limitations
- Architecture commitment: Letta wants to be the foundation your agents live in, not a light add-on
- Overkill if you just want to add memory to an existing agent (Mem0 is simpler)
Best for
- Teams building a system of persistent agents from scratch who want a real runtime
- Devs who want the agent to manage its own memory rather than re-injecting everything by hand