thread about memoria.day, an unconf geared towards nurturing a scene around memory systems - not so much even the tools, but the practices (tho also the tools) my main goal was to get a workflow going with LLM-generated flashcards & reviews—and I succeeded
Memory systems unconference and LLM-generated flashcards
16 tweets · September 2025 · 27 likes · 3 retweets · read on Twitter
first session was @michael_nielsen sharing some reflections that vibed with this article on "memory culture": better tools will come from scenes it included him sharing some of his own flashcards, which definitely felt like part of the memory culture vibe michaelnotebook.com/mmsw/
(I then encouraged the organizers to move the next big session to a medium-sized room not the big courtyard, so it wouldn't ruin the hallway track by taking over the dozen convos happening there & by sucking random others in. this helped—smart of them to have comp'd my ticket!)
that session was @kirkbyo & @andy_matuschak on "Why can LLMs win the IMO but can’t write good flashcards?" almost everybody was surprised by how bad LLMs are at this, despite lots of few-shot prompting. and I don't think we got to the bottom of why! notes.andymatuschak.org/zMyVLkT2M9DUrR…
my guess: there's some mix of instructions & 5-10-shot prompting that could get quite good cards to come out—that going from 1000 to 10k isn't gonna help if going from 100 to 1000 didn't produce any alpha the other key is to get ongoing feedback loops on which cards get kept
I had a few other good random conversations including one about how to journal until inner conflict resolves itself by being all brought into light. that process is generally bottlenecked on courage though it feels like it's bottlenecked on time.
as part of my "get a workflow", I made a point of getting @giacomo_ran to demo his app Rember, and was impressed that it managed to autogenerate good-enough flashcards for me to start remembering terms from The Unaccountability Machine. it was able to follow instructions!
many of these flashcards were trash, and some needed editing, but the point is that I could now have a quick "read chapter, review flashcards" workflow, rather than a "read chapter, make flashcards, review flashcards" workflow, which is too onerous so I simply do not. have not.
this highlights something interesting about different people's needs: Andy has written thousands of flashcards & already has a pipeline/process & doesn't want to look at bad ones. but for me, merely having the "get started" step take <5mins was most of the battle!
I now think that LLMs are gonna make a huge difference for the world of spaced repetition systems. in addition to cool new things like having the LLM ask you about the subject more conversationally, merely getting to skip "painstakingly write flashcards" will bring *millions*
and I think rember.com is already good enough! I told Giacomo to raise prices, so maybe get in now ;) (it has an anki plugin so you can send your cards there if you want, or review them inside rember on desktop or mobile web)
the way to do it for books is to use bookslicer to extract each chapter of a book you're reading, and then as you finish the chapter, toss the markdown file into rember and get it to turn it into cards. boom.
you could also do it as you start the next chapter: like the obvious trigger moment is when you read a jargon term and you're like "wait what was a 'resource bargain' again?" — you don't want to re-read the previous chapter. that means it's time to hit the flashcards!
there was a cool "homemade tools demo", during which I demo'd bookslicer, and @saulmunn demo'd an app called roadmapper that lets you easily make flashcards for streets in your city, and people showed off all sorts of things simple and crazy-complex jacobgw.com/road-mapper/
overall I feel excited by witnessing the dawn of a memory culture (even as NOT a power-user), and encouraged by this simple day of scene-building I've been thinking of doing some scene-building around memetics or evconch or organic alignment and it feels more in reach now
also partway through writing this thread I started vibecoding an upgrade to bookslicer so that it will now offer to let you download the book as a bunch of distinct files you can upload separately when the moment hits may also be good for the more general LLM use-case too
