Letta (formerly MemGPT) manages a virtual context window for LLMs.
| Feature | Rice | Letta |
|---|---|---|
| Context Management | Natural decay and attention-based drift | Explicit paging/swapping like OS RAM |
| Control | Developer/agent driven via API calls | LLM decides when to read/write memory |
| Complexity | High-level SDKs | Requires setting up a server runtime |
Rice uses natural decay and attention-based drift. Context flows organically based on relevance and recency. Letta explicitly pages blocks of text in and out of the context window, similar to how an operating system manages RAM.
Rice is developer/agent driven. You call focus(), drift(), commit(), and reminisce() to control memory. Letta is LLM-driven. The LLM itself decides when to read and write memory via function calls.
Rice provides high-level SDKs that abstract away the underlying complexity. Letta requires setting up a server that acts as the agent runtime.
Rice is practical for building defined agent behaviors with deterministic control over skills and natural context flow. Letta is powerful for infinite context simulations where the LLM manages its own memory.