Letta’s Memory Blocks and new Agent Framework

Letta (letta.com), featured in a recent deeplearning.ai course, extends MemGPT and is gearing up to become a key player in the agent framework space.

Memgpt was a foundational piece of work, making an analogy between LLM application memory (e.g, chat history) and OS virtual memory.

Letta encapsulates memory blocks to build agents now. The library API seems to be quite lean and appealing.

Features:

🌟 Memory divided into blocks. Block names are key to block value. It seems that there are no partial updates to value, only overwrites.

🌟 Memory get or update; build custom memories

🌟 Get Agents working by sending natural task commands to them

🌟 Agent converts commands to tool or memory calls.

We've seen a bunch of other Agent frameworks. How does Letta compare?

❗Crew agents don't explicitly take a memory object as a parameter

❗OpenAI's Swarm framework requires to build/manage own memory.

By making memory objects as first class citizens in their Agent framework, Letta enables a new set of fine-grained controls during multi-agent orchestration.

Also would be interesting to see if Letta's memory blocks could work seamlessly with OpenAI's Swarm.