Call with Tejas Channappa — 2026-06-07 · Transcript

Raw transcript of the call. Auto-transcribed via mlx_whisper (large-v3) on 2026-06-08.

Note on cleanup: The source audio (~88 min) was a ~15-minute substantive conversation followed by a long silent stretch. Whisper hallucinated a repeating "just, you know, just, you know…" loop over the silence, which has been stripped. A few residual phrase-duplications from the loop boundaries were also collapsed. No substantive content was removed — verified against the unique-line extraction of the raw file. Speaker labels are not available (single-channel, no diarization); the exchange is between Jacob and Tejas.

Structured extraction (ideas / follow-ups / open threads / questions) lives in tejas-2026-06-07.

Transcript

Go for it.

Yeah, I mean, there's a lot of, I guess, ideas around it. And I've been exploring some of it. Actually, I hope I could have showed one demo I did. It was pretty cool. But I think I broke it right now. So I guess, like, no, I literally have at least three or four prototypes that I've explored some ideas around it.

And what I figured was, yeah, we need something to ground us. And that something to ground us is just building the tool for ourselves and see how it helps us. Because I'm just spinning my wheels here.

Right, right, totally. And I can take this. And I can see it's super useful. But I can also see this meeting Outliner has to, obviously, be connected to your knowledge base. It's basically, at the end, what this ends up being — basically the Slack tool that I shared, or that Greg Eisenberg was talking about. So this becomes like an all-in-one Slack kind of a thing where you have meetings, you have conversations. And all of those is automatically maintaining your knowledge base for you.

And for maintaining that knowledge space, I think I have some opinions on that. And I found the missing thing. And the missing thing is like — the one main thing is, we do have to keep extracting entities from our discussions, right? Like raw transcripts and messages are just streams, raw streams. And then we extract entities from these streams with provenance. And these entities are the atomic knowledge of your base. Okay? And they all come with provenance based on where you have this conversation, or where this insight, question, decision, feedback all came from. So you're extracting different kinds of entities from the streams.

The missing piece is basically having a knowledge product output. So we got to tell the agents what we are interested in to develop this conversation. And for example, it could be the Nord Stream requirement documents, right? We just say the output that we were looking for, and we can have drafts. Like a section of the product could be just like, hey, these are all the knowledge products that we care about. And we can have, okay, Nord Stream requirements. And you could have specifications and whatnot. So that's going to anchor the agents to grab items from the raw streams, extract entities, and add patches to the right places — because they're anchored into "okay, this conversation actually relates to Nord Stream, and Nord Stream there's a knowledge product, which is a live dynamic product of requirements. So let me add this here, with the right problem."

So that's kind of like the bigger picture. But for meetings itself, for us, it's about augmenting short-term memory.

Yes, I love that phrase — augmenting short-term memory. Hashtag highlight.

Yeah. So it's about tracking all the different threads we run, and helping us move through these threads. And it could also have aspects of the HUD that you were talking about, right? So you can make use of the HUD as a feature to also grab relevant atoms from your knowledge base and say, "hey, actually, this might be relevant for this discussion." And that's one particular aspect, but the bare minimum is you, or whatever.

And like, that's kind of nice, because Claude is so good at HTML — it's so good at making amazingly presented information. Right? Exactly. And it's much richer than what you'd make on dashboards, like BI tools or whatever. Oh yeah. [It] has an agent. So it's kind of like Lovable for wikis.

Yeah, that's been top of mind for me. I mean, the point is, we get 15 years of building this library for taking notes, you know?

Wait, what is it? Is TiddlyWiki good for taking notes?

So the philosophy is really good — if you're in the right philosophy. And it's basically completely embedded in malleable software, because the whole philosophy, again, is — before agents, right? — the fact is, if you learn Tiddly, how to write Tiddlers, you have full control over your note-taking app.

Does it work on mobile very well?

I'm sure there's mobile versions of it. It's open source, right? So there's a TiddlyWiki which is not just a single HTML file — a TiddlyWiki which is good for mobile. TiddlyWiki which is really built for academic research and note-taking, with a lot of nice affordances for references, citations, provenance. So the idea is, if you can master TiddlyWiki, or writing Tiddlers, then there's no difference between you writing notes and changing the tool. It's the same language.

I think that went slightly over my head. I'm poking — you're on a TiddlyWiki at the same time, by the way.

Well, so you're saying that taking notes and changing the tool is not different. And I have a whole different take on that. But what is your take? And then I'll tell you my take, because I think it's important as well. It's a key insight that I wanted to share.

Yeah, I mean — for me, that felt like really an amazing philosophy for malleable software, right? Like, I was trying to use Obsidian for so long. And then I was like, I'm not sure if I'm going to be able to do that — try to build plugins and stuff, right? Before, it's just too much of a hassle to customize it. Or it's not that easy. Although Obsidian is still definitely somewhere on the spectrum of how easy you can modify the tool for your own needs. TiddlyWiki just takes it to another level. But it's also still something that you have to learn.

And actually, that's an amazing example — someone wrote a whole book called Grokking TiddlyWiki on TiddlyWiki, which teaches you how to use TiddlyWiki in TiddlyWiki with spaced repetition and all of those things. So it's insane, the things you can do with it. But again, it's another thing — now it's a skill. You need to learn it. So that was definitely a lot of friction.

But now, when you have agents, you don't have to worry about that, right? The language comes — the agents can write that Tiddly language. And you get the library for building notes. And the fact is, like HTML — we can literally keep talking, keep having conversation and build the tool as you're talking.

[Jacob:] I don't — not saying it's a better idea, but I want to share what I've been thinking, which is one step less hardcore than that. So you'll be able to screw around with it shortly. It's on our TestFlight. It's called Murmur — and that's just the name Claude gave it. It's an extremely minimalist thought stream that has a "send to agent" feature on every note that doesn't suck. I wouldn't test it till tomorrow because I'm pushing a bunch of stuff right now. But all it is, is you add a voice or text note at the top. And then you click the "send to agent" button, optionally. And that starts a chat with the agent based on the note. And the agent can bridge to my Mac mini or whatever — or my head[less] center box that I'm doing my cloud development on. Or, now it can bridge to my Hermes. I just stuck my Hermes on my Mac mini because that already has all of my everything.

Yeah.

[Recording continues into silence beyond this point.]