Love the structure of this article. Looking forwarw to slow reading it later today.
Based on this recent study from anthropic, it doesn't seem like that a more refined RAG system will not really solve the alignment problem or reduce the externalities of unpredictability from autonomous agents down to a manageable level that would make sense for a real world deployment just yet. Curious to hear your thoughts on this.
Hi Nir, thanks for this comprehensive article on managing conversational memory. It’s something I’m actually dealing with right now so this was very helpful.
One thing that always bugs me about similar articles I’ve read is when the author doesn’t describe the actual work that needs to go into each suggested method. I’m not talking about coding, I’m talking about categorizing your data. For instance, in a graph database, it doesn’t magically create the relationships for you, and sometimes once you exhaustively listed the classes or subdomains, it can take a lot of work to decide how you want them to be connected/related. As the tools evolve, they will be able to analyze your data and suggest both categories and relationships. But if you’re a medical specialist, or work for one as a data engineer, you usually already have the domain expertise, but building or updating the relationship graph can be tricky, including for things like avoiding redundancy and circular references.
Would you say that most of the techniques you mentioned are now mature enough to do most of this work, or we’re not there yet so developers and data scientists still have to do the work to define what aspects of their data they need to capture, relate, and recall and then test each configuration over time and make adjustments along the way?
Hi Nir thanks for the article. I have one question though, we are collating information over time mostly using summarization. This summarized info is from multiple documents. Now it can happen that one of the document is deleted. It is very expensive to re-summarize. Any methods which can help?
Olá, muito bom :). Obrigado.
Hello, Pure Gold. Thanks
You are welcome :))
It's just gold, thanks!
Thanks for the feedback! You are welcome
Most “AI agents” are still goldfish. Real autonomy needs hybrid memory: retrieval + compression + scoring, not just sliding windows.
Love the clarity of this article and simplicity of the examples. Makes a relatively technical topic easy to understand.
Thanks for that feedback! Helps me to know how to write the next ones :)
Amazing write up! Thank you
Thanks for the feedback! Happy you liked it :)
Love the structure of this article. Looking forwarw to slow reading it later today.
Based on this recent study from anthropic, it doesn't seem like that a more refined RAG system will not really solve the alignment problem or reduce the externalities of unpredictability from autonomous agents down to a manageable level that would make sense for a real world deployment just yet. Curious to hear your thoughts on this.
https://www.anthropic.com/research/project-vend-1
Hi Nir, thanks for this comprehensive article on managing conversational memory. It’s something I’m actually dealing with right now so this was very helpful.
One thing that always bugs me about similar articles I’ve read is when the author doesn’t describe the actual work that needs to go into each suggested method. I’m not talking about coding, I’m talking about categorizing your data. For instance, in a graph database, it doesn’t magically create the relationships for you, and sometimes once you exhaustively listed the classes or subdomains, it can take a lot of work to decide how you want them to be connected/related. As the tools evolve, they will be able to analyze your data and suggest both categories and relationships. But if you’re a medical specialist, or work for one as a data engineer, you usually already have the domain expertise, but building or updating the relationship graph can be tricky, including for things like avoiding redundancy and circular references.
Would you say that most of the techniques you mentioned are now mature enough to do most of this work, or we’re not there yet so developers and data scientists still have to do the work to define what aspects of their data they need to capture, relate, and recall and then test each configuration over time and make adjustments along the way?
Hi Nir thanks for the article. I have one question though, we are collating information over time mostly using summarization. This summarized info is from multiple documents. Now it can happen that one of the document is deleted. It is very expensive to re-summarize. Any methods which can help?