Comment by vessenes
12 hours ago
I’m on the other side of this - I’d like it to happen passively and in an automated fashion. Right now I’m playing around with porting concepts from zettelkasten (card based handwritten knowledge systems) to openclaw memory.
Rather than just coalescing to markdown files, the memory-zet plugin looks for actionable durable information and files it inside the existing zettelkasten system with embeddings - a quick no-LLM step (well 300m parameter query embed, it’s fast) is run against incoming chats or as a tool - this returns cards (zettels).
Zettels are somewhat unique in that the original methodology included a post-writing categorization and linking step - I have the system doing this as well. Result - cards can give you a (possibly cyclic) directed graph of connectivity. I built it for ‘centaur’ mode, so I can edit, link, unlink, move, etc through a nice little web interface.
The auto links are not the same quality I would make. But they are genuinely useful; upshot is for anything incoming, the LLM can see information directly about the query (if we have it), stuff that relates whether or not it embeds similarly, and can follow up links if they look promising with a fast tool call.
I made this memory system my daily driver yesterday; so far it is a significant improvement over the core memory extension (write to markdown files, don’t worry about compaction bro, it will be fine)!
It’s already building out people and organizational card bases for things that come in via email and whatsapp - this is a dream, basically. I think it will scale over time - but it’s at least scaling nicely over a few days of work right now.
You’re the first comment as I scrolled challenging the premise. But your solution amplifies what I see as a problem.
I’d like to add, that by forcing myself to look up the answer every time I have happy accidents where I learn new ways to do things.
It’s a skill to be willing to unlearn and always presume yourself ignorant, even if you do know how to do it. It’s like confirming “is my way still best practice.”
I'd love to hear more about this; have you considered creating a blog post, gist or something similar about it?
This sounds fascinating, if you ever write more about it I'd love to read it!
Oh man, it’s on the extensive backlog. :) I’m just about to put it up for people to play with, though, so you can clone it and use it or just ask claude to tell you about it.
I think the essay will be something like: adding structure post-hoc lets you build intelligence into the datastore as an architectural matter, not just rely on connections being made during use-time inference, using an embedding with links like this is much different than bulk embedding search, and we need some sort of tests to understand if this helps in practice, although it a) feels pretty good and b) it’s VERY nice to be able to refer to and modify the agents “mid term” memory directly in any event.
Anyway you’ve triggered me enough to say I’ll try and get the repo published today so people can look at it.
I’d love to see that too. thanks.