What are the best frameworks for building autonomous AI agents with long-term memory?

All answers to this question.

For long-term memory, LangGraph is currently superior to standard LangChain because it handles state as a first-class citizen. You should look into a "Checkpointer" pattern which saves the state of your agentic thread to a database like PostgreSQL or Redis. For the "knowledge" part of memory, Pinecone or Milvus are industry favorites. The trick is to separate "short-term context" (the current conversation) from "long-term summaries" (what the user liked last month). You can have a background agent that periodically summarizes past interactions to keep the context window clean.

   Answered 2024-09-12 by Amanda Collins

Are you planning to use a specific LLM like GPT-4o or Claude 3.5 Sonnet, or are you trying to keep the entire stack open-source with something like Llama 3?

   Answered 2024-09-14 by Jessica Reed

• Jessica, I'm actually leaning towards Llama 3 on-prem because of data privacy concerns in our industry. Does that change how I should handle the memory embeddings? I'm worried about the token cost and the embedding latency if the memory grows too large over the next few months of operation.

     Commented 2024-09-15 by Michael Stevens

You might want to check out MemGPT. It’s specifically designed to manage LLM memory as if it were an operating system, swapping data between RAM and disk.

   Answered 2024-09-16 by Christopher Brown

• Christopher is right. MemGPT is a fantastic mention for this specific problem. It really helps when you have a massive amount of historical data that simply won't fit in a standard prompt.

     Commented 2024-09-17 by Amanda Collins