[HN Gopher] AgenticMemory: Zettelkasten inspired agentic memory ...
       ___________________________________________________________________
        
       AgenticMemory: Zettelkasten inspired agentic memory system
        
       Author : simonpure
       Score  : 49 points
       Date   : 2025-03-03 18:09 UTC (4 hours ago)
        
 (HTM) web link (github.com)
 (TXT) w3m dump (github.com)
        
       | th0ma5 wrote:
       | Has a section for results that seems to just define what results
       | are. Other than mirroring what other software is doing, how do we
       | know any of these concepts are viable in the long term?
        
         | barbazoo wrote:
         | > 4.4 Empricial Results
         | 
         | https://arxiv.org/pdf/2502.12110
        
       | theodorewiles wrote:
       | yes i have been thinking about this for some time. One other
       | thing with zettle I'm not sure was implemented here is you can
       | have topic notes that just refer / summarize other notes; it
       | would be very interesting if these could be autonomously created
       | by some kind of clustering algorithm based on underlying links.
       | Kind of like summary-of-summary.
       | 
       | Also curious if there might be some improvements if you dont rely
       | on semantic similarity and just do all the pairwise "how related
       | are these memories and in what way" LLM test like
       | https://www.superagent.sh/blog/reag-reasoning-augmented-gene....
        
       | MrLeap wrote:
       | I've been waiting to see some paper that is like "a shallow tree
       | of key/values is all you need" to tackle model plasticity.
       | 
       | AI memory seems predominately a tension between compression and
       | lookup speed.
       | 
       | These large vectors are keys for lookups, a language of
       | expression, and a means of compression. Learning new things is
       | always easier when you can map it back to something you already
       | know. There's this page about forest fire simulations in a
       | scientific computing book I read back in college more than a
       | decade ago. I remember it viscerally because I've solved 100
       | different problems with that as its seed. I can barely remember
       | anything else in the book. I don't remember it because I read it
       | over and over, I remember it because it was useful and kept being
       | useful.
       | 
       | If some new technique or idea is 90% similar to something I
       | already know, I'll learn it easily. If it's 60%, I need to churn
       | it around, put in a lot of learning effort. If it's 0%, it's
       | noise from this angle.
       | 
       | > 4. Establishes meaningful links based on similarities > 5.
       | Enables dynamic memory evolution and updates
       | 
       | Wondering how much compression occurs in #5.
        
         | manmal wrote:
         | Fast forward and B+ trees end up making sense (again).
        
       | fudged71 wrote:
       | Coming from the TFT space of Roam/Tana/Obsidian... this is
       | awesome to see.
       | 
       | Further still, it would be neat to see a hybrid system where
       | humans and agents collaborate on building and maintaining a
       | knowledgebase.
        
       | fallinditch wrote:
       | This is interesting. Does this approach mean that it becomes
       | possible to use conversations with an LLM to fine tune the LLM in
       | very specific ways?
       | 
       | In other words, if this AgenticMemory can give structure to
       | unstructured conversations, and if this makes conversational
       | feedback more useful for the model to learn, then can we use it
       | to continually refine the model to be better at our particular
       | use case?
        
       | manmal wrote:
       | That's cosmic. I was wondering about the feasibility of exactly
       | this system today while driving to work. Zettelkasten seems
       | formalized/rule-based enough to work, and the limited context of
       | each Zettel should be ideal for LLMs.
       | 
       | I guess the biggest risk is that two related notes don't end up
       | getting connected, so the agent can get stuck in a local optimum.
       | I guess once a certain total number of notes has been reached, it
       | becomes quasi impossible to make all the connections, because
       | there are just too many possibilities?
        
       ___________________________________________________________________
       (page generated 2025-03-03 23:01 UTC)