i have been reading through the persistence / memo...
# koog-agentic-framework
g
i have been reading through the persistence / memory docs and was wondering if there are plans to make that process a little more automated. there are a bunch of frameworks out there that do this, but something like https://gibsonai.github.io/memori/, i found today could be interesting. i suppose this could be potentially wired into a koog agent externally. So it kind of begs the question of whether the intention is to have koog do things like this itself, or to delegate memory functions to other languages / frameworks, or both?
d
Memori is a framework for Python. How would you integrate that is your Kotlin-based agent?
Also, what features are you missing?
plus1 1
g
i'm not sure how i would integrate it without proxying my completions through a python server or a local process. what looks interesting about memori is that it claims to automate a lot of memory management. it seems there is some overlap with the persistence features in koog, but it goes further by deciding how / when to move memory types around between storage mechanisms etc. i might be missing something because i briefly read through the docs
v
There’s
nodeSaveToMemoryAutoDetectFacts
in Koog as well, but in my experience fully-automated approach doesn’t really scale well for larger tasks — it’s still better to list what you actually want to persist, and it works much more predictable. Usually, I’d just do the saving part at the end of your strategy, and loading — at the beginning (not of the whole history but of some important facts using the built-in fact retrieval mechanism).
move memory types around between storage mechanisms
Could you please elaborate a bit more what do you mean by “storage mechanisms” ?
g
I mean context vs a database vs vector storage
I don't have a lot of experience with memory in agents so I'll have to take your word that automated approaches don't work as well until I try building something fully functional in koog