Most agent frameworks treat the LLM as the product and memory as an afterthought — a vector store bolted on the side. Letta inverts this: the agent's memory is the system, and the LLM is a swappable component underneath. That single design choice is what lets the same agent keep learning about you across thousands of sessions instead of resetting to a stateless tool every time the context window fills up.
What Sets It Apart
- Self-editing memory from the MemGPT paper: the agent decides what to promote into its persistent core memory and what to page out, so personalization compounds rather than being re-prompted each turn.
- Model-agnostic by construction — memory and agent logic live outside the LLM, so you can switch providers or run local models without rewriting the agent.
- State lives on a server, not in your prompt: agents are addressable entities you deploy and query via API, which is a different mental model from stuffing history into every request.
Who It's For
Great fit if you're building agents meant to persist — personal assistants, support bots, or anything where the agent should remember a user weeks later — and you want memory as infrastructure rather than hand-rolled retrieval. Look elsewhere if you need a quick stateless completion call, a turnkey chat UI, or a no-code tool; this is an agent runtime that assumes you're writing the application around it.
