@Siv Darmalingum I can not find the tutorial I coaxed out of Gemini on this. It's in my chats somewhere. Not being able to find old subjects in my past chats, even though I copy/paste them into word files offline, is a big reason for my focus on memory right now. If you just sit down with your LLM, and describe what you want; a writing avatar it should be able do a back and forth Q/A with you. I did it more organically, a bit here and a bit there. It needs to be polished though. One of my first projects with LLMs was helping me write a series on the Jungian archetypes. The CEO of one of the groups I am working with is big time into them. I was having difficulty with them, and knew the LLMs could do personas. I asked it to talk to me like Carl Jung. It asked how detailed did I want it. In the back and forth, and the research of online resources, it came up with a very good persona. I just carried the idea forward. I've seen several videos which discuss and explain it on YT. That one just happens to be the one I watched.

@Siv Darmalingum no I don't mean do you use prompts collectively when working with your LLM. I mean specifically the summaries of the ongoing chat, which all of the LLMs will do if you ask it to. Chatgpt calls it a "Master Control Prompt" or MCP. You typically upload it and two other files when you start a chat, or continue on a chat which has become buggy because of length. I have an MCP for the overall project summary. That teaches the new version of the LLM what we are doing. In this case; I'm writing a book and the LLM is helping. Then I have a MCP which defines my preferences, boundaries and limits for the LLM's output. These tell it how I want it to work with me. Things like "No emojis". It also defines what kind of general persona I want it to interact with me with. Then I upload a third MCP on the properties of my "GrimDavid" writer's personal. This second persona allows me to use a voice which outputs to about 75% of ME and my style of writing. I toggle in and out of that voice as needed for generating chapter text. You should have at least the first two for all the subjects you are chatting in. The more specific you make the topic discussion, the tighter the output. You can also do separate MCPs for partial or for side chats. Most LLMs allow you to open a side chat, and tell it to not apply any of the data generated in the side chat, to the main subject in that chat. Tell the LLM the opening and the end point of the MCP in the chat, and the end point. Then you can take that, and open a brand new chat to discuss that subject alone. If you want, I can copy/paste one of the ones I have here, for you to see the formatting. Let me know.

Congratulations on the win, @Richard Clifton

I can't speak directly to why use 'folders' @Alejandro Morales because I'm just starting to look deeper into them as a system. I've been more focused on Nate Jone's Second Brain and was about to build one, then Karpathy released his "Knowledge Wiki" concept, and then last week, I came on the folder system. Now I'm re-evaluating options. But to answer your question, they are all ways to keep more of the benefits of YOUR interactions with AIs and LLMs under your control and making you money, not some hyperscaler selling you a monthly membership. Once they have locked YOUR data, and your chats, behind a paywall, then how much will you pay to access it? With personal offline memory, you control who uses it and if the membership or token cost goes up with current service, you take your toys and go somewhere else. - It's not just control of your data either. First it's all of the customization and boundaries, the preferred formats of outputs, the style of interactions with the LLM; all of that stuff is what turned you from a normal human output to the 5X human with AI assistant output. And hopefully, from a normal human paycheck to a paycheck 5x as big. What happens when that current company says, "Nah, we'll just keep your salary the same, even though your productivity is now much more". Do you quit? If you do, and you go to a new job interview, how do you show that YOU deserve that 5x paycheck? You do it with examples, with slide decks, with demos of any agent work (within the privacy and trade secrets of the previous business). Your AI memory is your new resume. - And finally, token prices, especially for the latest cutting edge LLMs is going to go up. Hyperscalers are subsidizing the big token users. That's going to end. Before it does, a lot of us are going to go with open LLMs hosted on our own local servers. Having most or all your data offline, means you run the simple stuff locally and only pay for the electricity, then only use the token stuff and higher end LLMs for the things that need it.

Congratulations on the win, @Roc Lee

@Carla Bosteder there were one or two interesting times <grins>

@Garth Frenzel new here as well. Welcome. If you don't mind me asking, is there a particular area of business you are focused on? A niche that has you excited? If its something you aren't quite ready to share publicly, that ok.