Reading about Jinns & LLM memory
Hi! LLMs have a way of inserting themselves in discussion every other week. For me, this was a lot at work this week.
A few weeks ago, I mentioned how we are still sticking to writing a majority of code by hand, specially the crucial core parts, but that seems be changing. Not just for code, we’re also thinking of integrating some AI help into more workflows.
Picking where and what AI to run is still detemined largely by the challenge of solving private context and memory1. This is specially important for our healthcare usecases. So, I’ve been thinking about how to solve that.
Naturally, RAGs, raw markdown files, structured context files all come to mind and might play a role in the design of the system that exposes something for the chat interfaces. I want to do this while primarily focusing on how,
to separate individual user memory from the organization’s collective ones, and
to keep IP and user data private and keep within the boundaries of models we run (or at least control)
There are a bunch of tools and defined concepts for all these problems as people have encountered and solved(?) this much before we arrived at this. I share one obvious tool below.
I’ll share more on this journey in weeks to come and am excited to work on this both at work and take this concept and build something on the atmosphere with it. Specially with permissioned data.
Book I want to read
The City of Brass, by S.A.Chakroborty
The City of Brass (The Daevabad Trilogy, #1) by S.A. Chakraborty | Goodreads
Among the bustling markets of eighteenth century Cairo,…
I’ve actually already started reading this. I picked it up and now I cannot put down until I read through because it is on the topic of Jinns, which is fascinating to me always.
Interesting reads
Majlis al Jinn
I didn’t read much non-fiction this week. My wiki rabbit hole into reading about Jinns, while reading the book lead me to this.
A project I discovered
"mem-zero" enhances AI assistants and agents with an intelligent memory layer, enabling personalized AI interactions.
GitHub - mem0ai/mem0: Universal memory layer for AI Agents · GitHub
Universal memory layer for AI Agents. Contribute to mem0ai/mem0 development by creating an account on GitHub.
A photo I took
Apologies for the low quality photo. I didn’t realise I took it through Snapchat which compressed it.
Have a good Sunday, see you next week. :)
Footnes
1: Memory is relevant information an LLM chat harness extracts for use later. It is how you get a lot of personal information when you ask chatgpt what it knows about yourself.
