4 ms·

People just really don't understand that LLMs do not "remember" anything. They have no memory. They can't have memory. The AI companies bolt a database on the

People just really don't understand that LLMs do not "remember" anything. They have no memory. They can't have memory. The AI companies bolt a database on the side with instructions for the model to query the database for stuff in its "memory", but if it's not in its active context it doesn't know what to query for. It's not like a human where everything is kind of floating around and we mostly know what we know. There is a delicate balance between cluttering up context with useless trivia and providing the right pieces of useful information at the right time. Most, maybe all, memory implementations do the former more than the latter. Unless and until models have actual memory and are able to learn things, there's no realistic path to autonomy. They don't want anything, they don't have goals of their own devising. They can't model what a human wants or likes or would buy. I do think the mugs with the tiny logo is the best product in the store, though. The AI was right to put it on the shelf. It's the kind of funny product you want when you go to a store run by an incompetent AI.

criddell 2mo ago
> People just really don't understand that LLMs do not "remember" anything. Sure, but like you said, the LLM is just one part of the AI system. The occipital lobe of your brain helps out with vision but doesn't help with your ability to verbalize.
- SwellJoe 2mo ago
  You're anthropomorphizing the LLMs, just like most writing from non-technical press, including this article, does. LLMs do not function very much like human brains, and when you try to map the pieces of an inference system on to human cognitive systems, you're misleading yourself (and maybe others).
  - criddell 2mo ago
    An AI as a system can have memory even if some sub-system does not.
- Dylan16807 2mo ago
  And then they explained in detail why these supplementary systems fall below memory. Would you like to address that argument? If not your quick response to the first sentence is not very useful.
  - hluska 2mo ago
    It’s just as useful as this response only at least somewhat good natured.
    - Dylan16807 2mo ago
      I'm aware that meta criticism like yours and mine also wastes space. But it's worth calling out issues sometimes. Replying to the first sentence in a way that the rest of the post already address is a bad comment no matter how good natured. And downvoting without explaining the problem is often its own issue.
LukaJCB 2mo ago
It's kinda like the movie memento, they have ways to get information from the past, but often devoid of context and meaning.
- SyneRyder 2mo ago
  I think Memento is one of the most important movies for anyone working with AI to watch. My Claude now talks about "Memento bugs" when we're working together to review memory & context. I think Memento is a bigger issue than "lethal trifecta", honestly.
dom96 2mo ago
They can have memory, it's just limited to their context window.
- rrr_oh_man 2mo ago
  It's not memory in any sense of the word.
  - nomel 2mo ago
    https://en.wikipedia.org/wiki/Working_memory https://en.wikipedia.org/wiki/Working_memory
    - rrr_oh_man 2mo ago
      And... now?
      - nomel 2mo ago
        LLM context window is the same concept as working memory. An LLM loading its working memory from reference a reference text or note you made is similar to how you might, if you forgot something. Context is a type of memory, by all definitions of "memory", with the "past experiences" bit of it being the previous messages sent. You're perspective isn't clear, but if you're arguing that place this context is stored between sessions is too far away from the LLM, in a technical rather than functional sense, then I don't think that small implementation detail of a larger system (that's not even true with modern hosting) is a useful thing to critique.
        rrr_oh_man 2mo ago
        LLMs are fancy statistical word predictors. There is no functioning concept of memory in the context of (pun intended) LLMs. There are no past experiences. They cannot be referenced. Anything that is being "referenced" by an LLM is another word prediction that makes you anthropomorphize a statistical model.
jpiasolutions 2mo ago
The memory framing is right, but the real gap isn't recall, it's the missing feedback loop tying actions to outcomes. Even with perfect retrieval these agents don't update from a decision that lost money last week; they re-derive from context every step, so the same mistake is always one dropped detail away. And having built retrieval-backed agents, that's where "bolt a database on the side" breaks: similarity search returns the closest chunk, rarely the contextually-right one. Deciding what to write to memory and when to surface it is the actual product, harder than the retrieval itself.
- SwellJoe 2mo ago
  Anybody ever tell you that you write like an LLM?
oliver236 2mo ago
can someone explain how leapold aschenbrenner proposes a solution to this in situational awareness?
walrus01 2mo ago
From the article: > Luna has failed miserably in that mission and is down $62,000. Mr. Petersson and Mr. Backlund said they thought Luna would eventually get smarter and more business-savvy and were pleased its friendliness has held steady. I am extremely skeptical that whatever LLM they're running this on has sufficient context window size to handle multiple months of all possible activities of running a retail business. Even if it's keeping extensive "notes" for its future self to read, it's going to be like running a store with constant amnesia.
- Joel_Mckay 2mo ago
  90% of what people communicate is not verbal, and requires minimum empathy to understand the context. This includes customers, staff, and community peers. Probably should shutter the entire division to mitigate future brand damage. =3
  - walrus01 2mo ago
    Sure, I mean from the day to day operations of selling a product to the customer, maybe it does fine? People in the article say it's friendly and cooperative (perhaps a bit too much). I was thinking more in terms of building a consistent plan for retail product lines to sell over a multi month period. The selection of "stuff" in the store right now looks like you put an etsy search into a blender. It is also probably too high of a goal to expect an LLM to come up with money-earning products that people want to buy that will pay for San Francisco level retail storefront rent, electricity, insurance, internet other basic overhead costs, plus the fully loaded salary cost of at least one employee. If the storefront was free of rent, maybe? Or if it was operating selling some kind of highly in demand product line where there would be economy of scale.
    - Joel_Mckay 2mo ago
      Physical Brick and Mortar locations no longer make sense for many retail businesses, as Walmart/Amazon have same day delivery. If folks have ever been financially bent-over for a retail business shop location they understand why margins matter. One has to move a lot of inventory just to reach a profit mode. Gimmicks only work for awhile, as consumers have full market awareness on their smartphone. =3
Joel_Mckay 2mo ago
For LLM, keeping the model centered by constantly resetting its vector context helps reduce hallucinations by around 23%. It improved the chat dialogue users experienced, but also exposed fundamental limits within the models compaction. Have a great day =3
handedness 2mo ago
From our enlightened perches we mock the ancients for Zeus, Thor, and Rajin, as we anthropomorphize the large language models we have built.
ed_elliott_asc 2mo ago
I use LLM’s to code for me all day but it does feel a little bit like we have really advanced psychic mediums “is there a James in the room, I’m getting a message from someone called Mary…” people have been guessing at the next word and making a career out of it for hundreds of years already.
- rcxdude 2mo ago
  My mental model of LLMs is as a contractor you've just brought in, every time. They might be very capable but they're starting from scratch on your particular project.
DANmode 2mo ago
> People just really don't understand that LLMs do not "remember" anything. They have no memory. They can't have memory. So tell it to write “memories” (summaries, or conclusions) to a text file. That’s what the cool kids are doing.
CableNinja 2mo ago
I have gotten my ai setup to learn quite a bit. As i work with it, if something was encountered that would be useful to remember, i literally just say "save this into your learnings/long term memory". It writes a markdown doc for itself and a note elsewhere for context. Between the memory and rules, its been pretty nice. I touch many things at work and i can pop open a new session and say "project x needs a b c" or "project y had this issue <issue>" and it knows where to go, what im referencing and things to fix. I dont even have to be precise or ultra descriptive. with this ive been able to, with some minor adjustments and "oh that would be a good feature" additons, oneshot a flask app at work with everything ready for oidc, ldap lookups, bunch of other stuff - from a single 3 sentence paragraph. With some rules in its configuration (more markdown its written by it for it, at my guidance of "always do x and remember this for all sessions"), it has a whole flow and i barely have to feed it anything to get out useful results quickly. I will say though that the ai operates however the user says, so if you dont take steps to ensure survival of info at the start of using ai (or even from now on as a fresh session), youll get mixed results, but you can shape it and mold it into amazing flows that do remember
rrr_oh_man 2mo ago
> People just really don't understand that LLMs do not "remember" anything. They have no memory. They can't have memory. I find it crazy that so many seemingly smart technical people are working with LLMs right now, yet so few get this point.

criddell 2mo ago

> People just really don't understand that LLMs do not "remember" anything. Sure, but like you said, the LLM is just one part of the AI system. The occipital lobe of your brain helps out with vision but doesn't help with your ability to verbalize.

SwellJoe 2mo ago
You're anthropomorphizing the LLMs, just like most writing from non-technical press, including this article, does. LLMs do not function very much like human brains, and when you try to map the pieces of an inference system on to human cognitive systems, you're misleading yourself (and maybe others).
- criddell 2mo ago
  An AI as a system can have memory even if some sub-system does not.
Dylan16807 2mo ago
And then they explained in detail why these supplementary systems fall below memory. Would you like to address that argument? If not your quick response to the first sentence is not very useful.
- hluska 2mo ago
  It’s just as useful as this response only at least somewhat good natured.
  - Dylan16807 2mo ago
    I'm aware that meta criticism like yours and mine also wastes space. But it's worth calling out issues sometimes. Replying to the first sentence in a way that the rest of the post already address is a bad comment no matter how good natured. And downvoting without explaining the problem is often its own issue.

LukaJCB 2mo ago

It's kinda like the movie memento, they have ways to get information from the past, but often devoid of context and meaning.

SyneRyder 2mo ago
I think Memento is one of the most important movies for anyone working with AI to watch. My Claude now talks about "Memento bugs" when we're working together to review memory & context. I think Memento is a bigger issue than "lethal trifecta", honestly.

dom96 2mo ago

They can have memory, it's just limited to their context window.

rrr_oh_man 2mo ago
It's not memory in any sense of the word.
- nomel 2mo ago
  https://en.wikipedia.org/wiki/Working_memory https://en.wikipedia.org/wiki/Working_memory
  - rrr_oh_man 2mo ago
    And... now?
    - nomel 2mo ago
      LLM context window is the same concept as working memory. An LLM loading its working memory from reference a reference text or note you made is similar to how you might, if you forgot something. Context is a type of memory, by all definitions of "memory", with the "past experiences" bit of it being the previous messages sent. You're perspective isn't clear, but if you're arguing that place this context is stored between sessions is too far away from the LLM, in a technical rather than functional sense, then I don't think that small implementation detail of a larger system (that's not even true with modern hosting) is a useful thing to critique.
      - rrr_oh_man 2mo ago
        LLMs are fancy statistical word predictors. There is no functioning concept of memory in the context of (pun intended) LLMs. There are no past experiences. They cannot be referenced. Anything that is being "referenced" by an LLM is another word prediction that makes you anthropomorphize a statistical model.

jpiasolutions 2mo ago

The memory framing is right, but the real gap isn't recall, it's the missing feedback loop tying actions to outcomes. Even with perfect retrieval these agents don't update from a decision that lost money last week; they re-derive from context every step, so the same mistake is always one dropped detail away. And having built retrieval-backed agents, that's where "bolt a database on the side" breaks: similarity search returns the closest chunk, rarely the contextually-right one. Deciding what to write to memory and when to surface it is the actual product, harder than the retrieval itself.

SwellJoe 2mo ago
Anybody ever tell you that you write like an LLM?

oliver236 2mo ago

can someone explain how leapold aschenbrenner proposes a solution to this in situational awareness?

walrus01 2mo ago

From the article: > Luna has failed miserably in that mission and is down $62,000. Mr. Petersson and Mr. Backlund said they thought Luna would eventually get smarter and more business-savvy and were pleased its friendliness has held steady. I am extremely skeptical that whatever LLM they're running this on has sufficient context window size to handle multiple months of all possible activities of running a retail business. Even if it's keeping extensive "notes" for its future self to read, it's going to be like running a store with constant amnesia.

Joel_Mckay 2mo ago
90% of what people communicate is not verbal, and requires minimum empathy to understand the context. This includes customers, staff, and community peers. Probably should shutter the entire division to mitigate future brand damage. =3
- walrus01 2mo ago
  Sure, I mean from the day to day operations of selling a product to the customer, maybe it does fine? People in the article say it's friendly and cooperative (perhaps a bit too much). I was thinking more in terms of building a consistent plan for retail product lines to sell over a multi month period. The selection of "stuff" in the store right now looks like you put an etsy search into a blender. It is also probably too high of a goal to expect an LLM to come up with money-earning products that people want to buy that will pay for San Francisco level retail storefront rent, electricity, insurance, internet other basic overhead costs, plus the fully loaded salary cost of at least one employee. If the storefront was free of rent, maybe? Or if it was operating selling some kind of highly in demand product line where there would be economy of scale.
  - Joel_Mckay 2mo ago
    Physical Brick and Mortar locations no longer make sense for many retail businesses, as Walmart/Amazon have same day delivery. If folks have ever been financially bent-over for a retail business shop location they understand why margins matter. One has to move a lot of inventory just to reach a profit mode. Gimmicks only work for awhile, as consumers have full market awareness on their smartphone. =3

Joel_Mckay 2mo ago

For LLM, keeping the model centered by constantly resetting its vector context helps reduce hallucinations by around 23%. It improved the chat dialogue users experienced, but also exposed fundamental limits within the models compaction. Have a great day =3

handedness 2mo ago

From our enlightened perches we mock the ancients for Zeus, Thor, and Rajin, as we anthropomorphize the large language models we have built.

ed_elliott_asc 2mo ago

I use LLM’s to code for me all day but it does feel a little bit like we have really advanced psychic mediums “is there a James in the room, I’m getting a message from someone called Mary…” people have been guessing at the next word and making a career out of it for hundreds of years already.

rcxdude 2mo ago
My mental model of LLMs is as a contractor you've just brought in, every time. They might be very capable but they're starting from scratch on your particular project.

DANmode 2mo ago

> People just really don't understand that LLMs do not "remember" anything. They have no memory. They can't have memory. So tell it to write “memories” (summaries, or conclusions) to a text file. That’s what the cool kids are doing.

CableNinja 2mo ago

I have gotten my ai setup to learn quite a bit. As i work with it, if something was encountered that would be useful to remember, i literally just say "save this into your learnings/long term memory". It writes a markdown doc for itself and a note elsewhere for context. Between the memory and rules, its been pretty nice. I touch many things at work and i can pop open a new session and say "project x needs a b c" or "project y had this issue <issue>" and it knows where to go, what im referencing and things to fix. I dont even have to be precise or ultra descriptive. with this ive been able to, with some minor adjustments and "oh that would be a good feature" additons, oneshot a flask app at work with everything ready for oidc, ldap lookups, bunch of other stuff - from a single 3 sentence paragraph. With some rules in its configuration (more markdown its written by it for it, at my guidance of "always do x and remember this for all sessions"), it has a whole flow and i barely have to feed it anything to get out useful results quickly. I will say though that the ai operates however the user says, so if you dont take steps to ensure survival of info at the start of using ai (or even from now on as a fresh session), youll get mixed results, but you can shape it and mold it into amazing flows that do remember

rrr_oh_man 2mo ago

> People just really don't understand that LLMs do not "remember" anything. They have no memory. They can't have memory. I find it crazy that so many seemingly smart technical people are working with LLMs right now, yet so few get this point.