Post

Let's start with what memory *is*: fundamentally, it's a capacity that permits continuous improvement. Memory makes it... * ..harder to repeat past mistakes * ..easier to reuse past solutions * ..cheap to derive new value from already expended effort (i.e. it amortizes computation)

Philosophical aside: every agent -- even you -- can be decomposed into a system comprised of interacting subagents. My behavior is the manifestation of the different conflicting facets of my psyche negotiating compromises.

Through this lens, any agent or project can be viewed as a system/institution. Similarly, any discussion of "memory" can be reframed in the context of the "institutional memory" of some system.

I think you're overestimating the complexity of what my caveman brain produces 🤪

This is good stuff. Ive been tinkering with LLM wiki patterns and the need for opinionated memory into shared stuff is quite apparent. I am bemused at how weak the coding genies default memory handling is, when there is rather a lot of prior art on what matters for getting work done.

Interesting. I have my own private agent note taking approach that works well, but yeah project level is hard. Some folks on the team like throwing everything into them, might be useful. My view is it is harmful context poisoning to share unpursued design ideas etc.

fair take. It's been my experience that it can be powerful/important to cite unpursued/unendorsed ideasas justifications though to keep those ideas from being revisited/relitigated. Otherwise it can be hard to point at something and say "yeah we tried that already and it didn't work, don't do that."

This way, I don't need to correct the LLM: it corrects itself. It might take a few extra tokens getting there, but it also necessitates less direct oversight from me because it does a better job keeping itself on the rails.