Your Agent Already Wrote It Down: All your Claude Code (or Codex) Sessions as Queryable Memory

Speaker 0: Hey. AI. Speaker 1: Okay. Speaker 0: Yeah. Speaker 1: Agent memory is a big thing. It's an odd topic. There's a lot of stuff going on. I built a simple system for harvesting all of my clogged coding sessions into an LLM Wiki. I have a lot of machines. Speaker 1: I have 5 different computers. I've run a lot of different projects for, a lot of different systems and AI wanted to get it all into 1 thing. SDE what's an LLL Wiki, is it's, this, August by Andre Carpathi. It Laes, you know, it's basically the the core idea is that you build instead of query. So a a Track system would query your corpus, but rather than that, an LMWiki will take your corpus and build a Date out of City. Speaker 1: And then it will maintain it. It'll do the diligence that, a human would never be willing to maintain of constantly updating your wiki. And, so thing number 2 is that if you were to look in a particular directory, your Twiddle dot quad projects directory, you would see a metric ton of JSONL files Andrey those are your sections. And, if you were to have a spot to tell you what the the schema is of those JSON elements. Like, okay. Speaker 1: There's these envelopes and, like, there's different types of envelopes. Some are users, some are, assistants, and they have different content blocks and some are text and tool use and then some looks very familiar. This is a trace. Right? So it's basically just a trace of your of your session. Speaker 1: And, so those 2 things together, you know, the the Provide just launched 1000 repos Sayantan of them was, this 1 by Solutions. And thank you, Pratayush, if you're out there watching this. And he they, created a system that, you know, harvests off of your, your Singh, user HINTS, another 1 wiki. And, so what I built AWS, a AI. And it's not especially ThetaEdge, but AI these are all my computers over here. Speaker 1: And I sync the files over SSH from Kansy. And, you know, there's this config file that I if I want to add a new machine, I just plug a new line in there. And then it gets over here and I harvest some every 6 hours, then I sync them. And that syncing is, just taking the JSONL and converting it to Markdown, into these raw sessions, which are the immutable transcripts. The LLM Wiki has this notion of the immutable transcripts, and then you have the actual Wiki with all the cross links and entities and all that. Speaker 1: And then you have the token attendees portion, the LLM ingest, which actually takes these new ones and then updates the Wiki, you know, and, then you get a AI, static HTML site which you can get with your browser and you get an 2026 server and then, yeah, you can just grep over the sessions in wiki. And, so this is the visualization of this is the website, and you get some, you know, cool, metrics, top line metrics, etcetera stuff. And you get your projects. And, you can drill into those and you can see the individual sessions. And you can do, you know, the typical filter search type of thing. Speaker 1: And then you have the more interesting about, like, your Kennedy. Andrey, like, for example, this is gonna be Alex Kochanik. He's 1 of my coworkers. And, 1 of the key facts about Alex is that he questioned the justification for the parent type field. That's faster. Speaker 1: He, so how have I built this thing and how have I used it? So 1 thing I did was I had built this other project that involves fine tuning, and then, deploying into production this, 7,000,000,000 perimeter model. And, so I asked, the Date to, you know, write me a Cloud Architect. And, the key thing here it's pretty good artifact. I like it. Speaker 1: The key thing here is they got all of this from the session data. It didn't actually look at my repo at all. It didn't so and then, I also another thing you can do with it is you can generate the proposal for your AI takers demo. So this is exactly the form that you need to fill out. And, I did that by having it AI the file on Wiring repo itself, very self referential. Speaker 1: Another thing you can do is discover that you've got some apps Down processes running around various machines in your network, and then you need to, like, shut them down because they're just doing, like what is it even doing? It's look at this is a system prompt or something. I don't know. Crazy. Another thing you can do is you can generate ADRs. Speaker 1: Andrey then the last thing I'll show is this was just a Singh that I ThetaEdge to get the design docs associated with a particular project at a particular point in time. And, so I Founder it marvelously useful and I continue to improve it as I go. Speaker 0: We've got time for a couple of questions. I'm gonna bring Lan up to plug in her computer now SDE that we Code have a smooth switcher over there. But who's got a question? Yes? Speaker 1: Yeah. So it's very clear how Speaker 0: that's usable from a a human perspective. But how like, for agents, is it significantly better than just record and session files? Speaker 1: Yeah. Great question. Here. So that was, that was a big question for me and I felt like I needed to try and understand it a little deeper here. So I started by just doing some analytics on the Sessions that I had, you know, the the users I put through. Speaker 1: And, AI my 1st theory was AI, oh, it's it's MCP, it's the MCP Tools. That's the great thing. Right? And so then I discovered that AI for every 1 MCP call, there were 10 inRepped calls. Andrey that's just because Cloud code loves City some graphs. Speaker 1: Like, clock code is old school. So then my next theory was, oh, built, you Speaker 0: know, Speaker 1: query grepping the the wiki is so much better than grepping the, Ben the Sessions the raw section transcripts. And then so I ran a test on that, and it turns out that, actually, no. If you're looking for specific things, it isn't better. So I did it, like, an eval and and a, you know, double blind study and and no. It isn't better. Speaker 1: The turn 2 more teach later. This is AI 10PM last night and AI, like, sweating my demo. And I I got AI I got to the bottom of City. And what it is is it the Singh is much much better for asking about things that you it knows. So if you wanna find something, then you it's best to Graph the trans the, the transcripts correctly. Speaker 1: But if you wanna know something, if you wanna, like, characterize, analyze, you know, look across times and, you know, things that it doesn't know at the point at which it new had the SSH, but later on it contacts her, then the New is far, far better and and much more it's more efficient because it generates fewer tool calls that solves more problems. We have time for 1 more question. New. Speaker 0: There AWS system prompts or instructions, I assume, for the various LLM Wiki processes. Did you, like, take some off the shelf, or did you write them yourself? It's Like, you gotta Guo, like, you gotta double check your citations or, like, summarize new documents. There's, like, prompts to tell, like, what it's doing at that time. Speaker 1: You know, I I used my my coding buddy. August took that built and ran with City. And, you know, Ben generated something that seems pretty awesome and Works well enough. Tuned it a little bit. Yeah. Speaker 0: Cool. Thanks, John.