i was building an agent to do outreach, but while i was leaning how basic memory works with letta - i noticed something; all memory architecture (letta/claw/claude) simply moves the pain to retrieval.
I started looking back at my own memories trying to understand how my mind remembers stuff as opposed to how agents do, and then i came across a pg essay that mentioned something similar.
Human memory doesn't store images in full resolution; just think back to any old memory you have - it's barely visible (100*100res). The older the memory is, the lower the resolution is. Faces are usually blurred; the images are there for context not for clarity. Same with text, when i look back at an old conversation, i don't remember most of the stuff that was said nor do i remember a "summary" of what happened. My memory of conversations is a lot like my memory of images; one part is ultra clear while the other parts are blurred (pain=weights).
Why would our minds blur stuff if they are saved to long-term memory!
if my 10-year-old laptop has 500gb storage, my brain definitely has many terabits if not x100 that!
Why do i have to relearn a skill that i used to be very good for years when i clearly remember almost all of the steps to how to do it. This sounds a lot like a context memory that went through compaction & then dumped to the disk. Even tho the log has been saved to long term memory; it's no longer usable unless rebuilt into a new context again, which is why i need to relearn the skill.
ONE POSSIBLE EXPLAINATION why the brain needs to care about memory saving to this a degree that it has to forget all of the details about my first car except its color & brand. It's that our biological memory is operating more like a set of context windows that switch between each other not retrieving data from a permanent long-tern storage!
Which means that we switch back and forth between different specialized context windows while we work (which is why we can switch takes in less than 60sec). Unless the context window is dumped into the long-term memory, we can still load it fast. The brain doesn't dump current sessions unless something major happened, when we change career, cities or we go through something big.
Whenever i am in a new city it feels like a new context window has loaded into my brain, my childhood memories suddenly feel different, the stuff relating to a different career i had in the old city feel alien. The new context window even makes me think in a different lens & analyze stuff from a totally different angle.
Why are we not doing the same with Ai, why do all memory structures either remove the full old sessions entirely or summarize them into oblivion? why not do what the brain is already doing & spread the work across 100s of context windows that each is concerned with 1 thinking pattern?!
I started looking back at my own memories trying to understand how my mind remembers stuff as opposed to how agents do, and then i came across a pg essay that mentioned something similar.
Human memory doesn't store images in full resolution; just think back to any old memory you have - it's barely visible (100*100res). The older the memory is, the lower the resolution is. Faces are usually blurred; the images are there for context not for clarity. Same with text, when i look back at an old conversation, i don't remember most of the stuff that was said nor do i remember a "summary" of what happened. My memory of conversations is a lot like my memory of images; one part is ultra clear while the other parts are blurred (pain=weights).
Why would our minds blur stuff if they are saved to long-term memory! if my 10-year-old laptop has 500gb storage, my brain definitely has many terabits if not x100 that! Why do i have to relearn a skill that i used to be very good for years when i clearly remember almost all of the steps to how to do it. This sounds a lot like a context memory that went through compaction & then dumped to the disk. Even tho the log has been saved to long term memory; it's no longer usable unless rebuilt into a new context again, which is why i need to relearn the skill.
ONE POSSIBLE EXPLAINATION why the brain needs to care about memory saving to this a degree that it has to forget all of the details about my first car except its color & brand. It's that our biological memory is operating more like a set of context windows that switch between each other not retrieving data from a permanent long-tern storage!
Which means that we switch back and forth between different specialized context windows while we work (which is why we can switch takes in less than 60sec). Unless the context window is dumped into the long-term memory, we can still load it fast. The brain doesn't dump current sessions unless something major happened, when we change career, cities or we go through something big.
Whenever i am in a new city it feels like a new context window has loaded into my brain, my childhood memories suddenly feel different, the stuff relating to a different career i had in the old city feel alien. The new context window even makes me think in a different lens & analyze stuff from a totally different angle.
Why are we not doing the same with Ai, why do all memory structures either remove the full old sessions entirely or summarize them into oblivion? why not do what the brain is already doing & spread the work across 100s of context windows that each is concerned with 1 thinking pattern?!