memo lines as attention signals for AI distillation

[jshph]

Been following Char for a while — love that you already do the memo + transcript → combined summary thing. That's the right call imo, it addresses the core problem where AI summaries have no idea what actually mattered to you in the call.

I've been building something in a similar space (aside — Rust TUI + local whisper, very different form factor) and took the memo idea a step further: each line you type gets a created_at timestamp, and if you go back and edit a line mid-meeting it also gets an edited_at. During distillation, edited lines get weighted higher — the reasoning being if you came back to revise something, that topic was important enough to reconsider.

The other thing I tried: instead of summarizing in isolation, the distillation step searches the user's Obsidian vault for related notes and writes the meeting note back with [[wikilinks]] to existing thinking. Ran it on a couple calls and it pulled in a half-finished doc I'd forgotten about, connected to what we were discussing. Not as an action item, just as context.

The whole distillation pipeline is a 240-line markdown file the LLM follows — not compiled into the app. You can read it, fork it, change how it prioritizes topics. Curious if that kind of transparency matters to people here or if the template/plugin system covers it.

Saw @dkindlund's point in #1978 about passing older transcriptions of the same meeting type for historical context — that's basically what I'm trying to do but with the whole vault instead of just past meetings of the same type.

Anyway — does anyone here actually connect their Char notes back to their broader note system? Or is the standalone summary good enough for most workflows?