Thanks Simon. I think this might solve one of the most common questions people ask me: how do I get Perplexity-like inline citations on my LLM output?

This looks like model fine tuning rather than after the fact pseudo justification. Do you agree?

Yeah, I think they fine-tuned their model to be better at the pattern where you output citations that reference exact strings from the input. Previously that's been a prompting trick, e.g. here: https://mattyyeung.github.io/deterministic-quoting

Makes sense. I wonder if it affects the model output performance (sans quotes), as I could imagine that splitting up the model output to add the quotes could cause it to lose attention on what it was saying.

As someone who interned at Palm, love this so much!

Oh wow, do you have any interesting stories from back then? Saw some unreleased prototypes?

Depends on the language/standard library. For example in C if your library includes its own HTTP library that's probably not a plus.

Congrats Taranjeet and Deshraj!

So after using Mem0 a bit for a hackathon project, I have sort of two thoughts: 1. Memory is extremely useful and almost a requirement when it comes to building next level agents and Mem0 is probably the best designed/easiest way to get there. 2. I think the interface between structured and unstructured memory still needs some thinking.

What I mean by that is when I look at the memory feature of OpenAI it's obviously completely unstructured, free form text, and that makes sense when it's a general use product.

At the same time, when I'm thinking about more vertical specific use cases up until now, there are very specific things generally that we want to remember about our customers (for example, for advertising, age range, location, etc.) However, as the use of LLMs in chatbots increases, we may want to also remember less structured details.

So the killer app here would be something that can remember and synthesize both structured and unstructured information about the user in a way that's natural for a developer.

I think the graph integration is a step in this direction but still more on the unstructured side for now. Look forward to seeing how it develops.

Thanks yding! Definitely agree with the feedback here. We have seen similar things when talking to developers where they want:

- Control over what to remember/forget - Ability to set how detailed memories should be (some want more detailed vs less detailed) - Different structure of the memories based on the use case

The short answer is he works at Cohere. But longer answer is that the model probably doesn’t matter that much.

Looks great! thanks for sharing your architecture choices here.

Congrats on the launch!

Very cool!