Context window management has become a critical part of every LLM application - from the basics (embeddings models, vector DBs) to more advanced techniques (query rewriting, HyDE, summarization). MemGPT is a new tool from UC Berkeley built by Charles Packer that automates "memory" management for LLMs and creates a functionally infinite context window. Charles joins us this week to talk about MemGPT, the techniques behind it, and where the conversational AI space is headed.
MemGPT: github.com/cpacker/memgpt
Charles' Twitter: / charlespacker
0:00 Intro
0:32 From AVs to conversational AI
1:55 Creative content generation with LLMs
4:53 The motivation behind MemGPT
8:19 When long context windows are enough
10:10 How MemGPT achieves longer contexts
14:38 How MemGPT relates to RAG
17:39 How MemGPT leverages function calling
21:20 Opportunities for OSS models to improve
22:59 Limitations of MemGPT
26:12 Future directions in conversational AI
28:37 Conclusion
Негізгі бет Ғылым және технология Generating Conversation: MemGPT, Memory Management for LLMs - Charles Packer (Episode 9)
Пікірлер: 17