Well here's the interesting thing to think about for me.

Human memory is.... insanely bad.

We record only the tiniest subset of our experiences, and those memories are heavily colored by our emotional states at the time and our pre-existing conceptions, and a lot of memories change or disappear over time.

Generally speaking even in the best case most of our memories tend to be more like checksums than JPGs. You probably can't name more than a few of the people you went to school with. But, if I showed you a list of people you went to school with, you'd probably look at each name and be like "yeah! OK! I remember that now!"

So.

It's interesting to think about what kind of "bar" AGI would really need to clear w.r.t. memories, if the goal is to be (at least) on par with human intelligence.

Memory is a skill- its plastic, not static.

You can get better at remembering things, like you can get better at dancing or doing exercise.

We can also specialize our memory to be good at some things over others.

Insanely bad compared to what else in the animal kingdom? We are tool users. We use tools, like language, and writing, and technology like audio/video recording to farm out the difficulties we have with memory to things that can store memory and retrieve them.

Computers are just stored information that processes.

We are the miners and creators of that information. The fact that a computer can do some things better than we can is not a testament to how terrible we are but rather how great we are that we can invent things that are better than us at specific tasks.

We made the atlatl and threw spears across the plains. We made the bow and arrow and stabbed things very far away. We made the whip and broke the sound barrier.

Shitting on humans is an insult your your ancestors. Fuck you. Be proud. If we invent a new thing that can do what we do better it only exists because of us.

Insanely bad compared to books or other permanent records. The human memory system did not evolve to be an accurate record of the past. It evolved to keep us alive by remembering dangerous things.

Books and other permanent records of human thought are part of the human memory system. Has been for millennia. If you include oral tradition, which is less precise, but collectively much more precise than any individual thought or memory, it goes much further.

We are fundamentally storytelling creatures, because it is a massive boost to our individual capabilities.

When I say, "Insanely bad compared to what else in the animal kingdom?" and you respond with, "compared to books or other permanent records"

"Books or permanent records" are not in the animal kingdom.

Apples to Apples we are the best or so very nearly the best in every category of intelligence on the planet IN THE ANIMAL KINGDOM that when in one specific test another animal beats a human the gap is barely measurable.

How do you know we have better memory than other animals?

This crap tier article was the first and easiest response to your question:

https://sciencesensei.com/24-animals-with-memory-abilities-t...

3 primate species where very concise tests showed that they were close to or occasionally slightly better than humans in specifically rigged short term memory tests (after being trained and put up against humans going in blind).

I've never heard of any test showing an animal to be significantly more intelligent than humans in any measure that we have come up with to measure intelligence by.

That being said, I believe it is possible that some animals are either close enough to us that they deserve to be called sentient, and I believe it is possible that other creatures on this planet have levels of intelligence in specialized areas that humans can never hope to approach unaided by tools, but as far as broad range intelligence, I think we're this planets' possibly undeserved leaders.

Can you find anything that I didn't consider?

I don't think working memory has much at all to do with sentience.

The conversation was more about long-term memory, which has not been sufficiently studied in animals (nor am I certain it can be effectively studied at all).

Even then I don't think there is a clear relationship between long-term memory and sentience either.

And yet I have vivid memories of many situations that weren't dangerous in the slightest, and essentially verbatim recall of a lot of useless information e.g. quotes from my favorite books and movies.

I am not sure exactly what point you're trying to make, but I do think it's reductive at best to describe memory as a tool for avoiding/escaping danger, and misguided to evaluate it in the frame of verbatim recall of large volumes of information.

Chimpanzees have much better short term memories than humans do. If you test them with digits 1-9 sequentially flashed on a screen, they're able to reproduce the digits with lower loss than undergraduate human students.

https://link.springer.com/article/10.1007/s10071-008-0206-8

> While the between-species performance difference they report is apparent in their data, so too is a large difference in practice on their task: Ayumu had many sessions of practice on their task before terminal performances were measured; their human subjects had none. The present report shows that when two humans are given practice in the Inoue and Matsuzawa (2007) memory task, their accuracy levels match those of Ayumu.

Hmm.

So? If I write something down as a child and forget it I can come back 60 years later and know what I wrote down.

Chimpanzees can not.

The question was whether there are animals who have better memory than humans. I named one: humans are not superior to animals in all cognitive capabilities.

See Nathan's response. They trained the chimp and threw the humans in blind against them.

Like I said, so close as to be almost immeasurable.

That's a very anthropocentric view. Technology isn't a series of deliberate inventions by us, but an autonomous, self-organizing process. The development of a spear, a bow, or a computer is an evolutionary step in a chain of technological solutions that use humans as their temporary biological medium. The human brain is not the starting point or center of this process. It is itself a product of biological evolution, a temporary information-processing system. Its limitations such as imperfect memory, are simply constraints of its biological origin. The tools we develop, from writing to digital storage are not just supplements to human ability, but the next stage in a system that is moving beyond its biological origins to find more efficient non-biological forms of information storage and processing. Human pride in creation is a misinterpretation. We are not the masters of technology. We're just the vehicle of it. Part of a larger process of technological self-improvement that is now moving towards an era where it might no longer require us

I think your understanding of the words "autonomous" and "self-organizing" is somewhat lacking. If there were no humans, those things would not happen.

Further, if it were a byproduct of the presence of humans, then the backpath of invention would be repeated multiple times and spread out across human history, but, for instance, despite the presence of saltpeter, sulfur, and charcoal, magnetite, wood and ink across the planet, the compass, gunpowder, papermaking and printing were essentially exclusively invented in China and only spread to Europe through trade.

The absence of the four great inventions of china in the Americas heavily implies that technology is not a self-organizing process but rather a consequence of human need and opportunity meeting at cross ends.

For instance, they had the wheel in America, but no plow animals, so the idea was relegated to toys despite wheelbarrows being a potentially useful use for the wheel.

My mental model is a bit different:

Context -> Attention Span

Model weights/Inference -> System 1 thinking (intuition)

Computer memory (files) -> Long term memory

Chain of thought/Reasoning -> System 2 thinking

Prompts/Tool Output -> Sensing

Tool Use -> Actuation

The system 2 thinking performance is heavily dependent on the system 1 having the right intuitive models for effective problem solving via tool use. Tools are also what load long term memories into attention.

Very cool, good way to think about it. I wouldn’t be surprised if non-AGI LLMs help write the code to augment themselves into AGI.

The unreasonable effectiveness of deep learning was a surprise. We don’t know what the future surprises will be.

I like this mental model. Orchestration / Agents and using smaller models to determine the ideal tool input and check the output starts to look like delegation.

The long term memory is in the training. The short term memory is in the context window.

The comparison misses the mark: unlike humans, LLMs don't consolidate short-term memory into long-term memory over time.

That is easily fixed, ask it to summarize it's learnings, store it somewhere, and make it searchable through vector indexes. An LLM is part of a bigger system that needs not just a model, but context and long term memory. Just like human needs to write things down.

LLMs are actually pretty good at creating knowledge: if you give it a trial and error feedback loop it can figure things out, and then summarize the learnings and store it in long term memory (markdown, RAG, etc).

You’re making the assumption that there’s one, and only one, objective summarization, this is entirely different than “writing things down.”

Why do you assume i assume that?

My bad if I misunderstood. I assumed by your use of “it” and approximation methods.

This runs into the limitation that nobody has RL'd the models to do this really well.

Over time though, presumably LLM output is going into the training data of later LLMs. So in a way that's being consolidated into the long-term memory - not necessarily with positive results, but depending on how it's curated it might be.

> presumably LLM output is going into the training data of later LLMs

The LLM vendors go to great lengths to assure their paying customers that this will not be the case. Yes, LLMs will ingest more LLM-generated slop from the public Internet. But as businesses integrate LLMs, a rising percentage of their outputs will not be included in training sets.

The LLM vendors aren't exactly the most trustworthy on this, but regardless of that, there's still lots of free-tier users who are definitely contributing back into the next generation of models.

For sure, although I'm fairly certain there is a difference in kind between the outputs of free and paid users (and then again to API usage).

Please describe these "great lengths". They allowing customer audits now?

The first law of Silicon Valley is "Fake it till you make it", with the vast majority never making it past the "Fake it" stage. Whatever the truth may be, it's a safe bet that what they've said verbally is a lie that will likely have little consequence even if exposed.

> great lengths to assure

is not incompatible with

> "Fake it till you make it"

I don't know where they land, but they are definitely telling people they are not using their outputs to train. If they are, it's not clear how big of a scandal would result. I personally think it would be bad, but I clearly overindex on privacy & thought the news of ChatGPT chats being indexed by Google would be a bigger scandal.

You did hear that it did happen (however briefly) though, yeah?

https://techcrunch.com/2025/07/31/your-public-chatgpt-querie...

That's my point. It is a thing that is known and obviously a big negative, but yet failed to leave a lasting mark of any kind.

Ah, the eternal internal corporate search problem.

That's only if you opt out.

ChatGPT training is (advertised as) off by default for their plans above the prosumer level, Team & Enterprise. API results are similarly advertised as not being used for training by default.

Anthropic policies are more restrictive, saying they do not use customer data for training.

Is this not a tool that could be readily implemented and refined?

my knowledge graph mcp disagrees

I think it's more analogous to "intuition", and the text LLMs provide are the equivalent of "my gut tells me".

Humans have the ability to quickly pass things from short term to long term memory and vice versa, though. This sort of seamlessness is currently missing from LLMs.

No, it’s not in the training. Human memories are stored via electromagnetic frequencies controlled by microtubules. They’re not doing anything close to that in AI.

And LLM memories are stored in an electrical charge trapped in a floating gate transistor (or as magnetization of a ferromagnetic region on an alloy platter).

Or they write CLAUDE.md files. Whatever you want to call it.

That was my point, they’re stored in a totally different way. And that matters because being stored in microtubules infers quantum entanglement throughout the brain.

Whether QE is a mechanism in the brain still seems up for debate from the quick literature review I tried, but would love to learn more.

Given the pace of quantum computing it doesn’t seem out of the realm of possibility to “wire up” to LLMs in a couple years.

are ANN memories not also stored in loops like recurrent nets?

It's not that either.

I don't believe this has been really proved yet.

There are many folks working on this, I think at the end of the day the long term memory is an application level concern. The definition of what information to capture is largely dependent on use case.

Shameless plug for my project, which focuses on reminders and personal memory: elroy.bot

But other projects include Letta, mem0, and Zep

What is the current hypothesis on if the context windows would be substantially larger, what would this enable LLMs to do that is beyond capabilities of current models (other than the obvious the now getting forgetful/confused when you’ve exhausted the context)?

I mean, not getting confused / forgetful is a pretty big one!

I think one thing it does is help you get rid of the UX where you have to manage a bunch of distinct chats. I think that pattern is not long for this world - current models are perfectly capable of realizing when the subject of a conversation has changed

I wonder if there will be some sort of bitter lesson, generalized memory beating specialized memory.

Yeah to some degree that's already happened. Anecdotally I hear giving your whole iMessage history to Gemini results in pretty reasonable results, in terms of the AI understanding who the people in your life are (whether doing so is an overall good idea or not).

I think there is some degree of curation that remains necessary though, even if context windows are very large I think you will get poor results if you spew a bunch of junk into context. I think this curation is basically what people are referring to when they talk about Context Engineering.

I've got no evidence but vibes, but in the long run I think it's still going to be worth implementing curation / more deliberate recall. Partially because I think we'll ultimately land on on-device LLM's being the norm - I think that's going to have a major speed / privacy advantage. If I can make an application work smoothly with a smaller, on device model, that's going to be pretty compelling vs a large context window frontier model.

Of course, even in that scenario, maybe we get an on device model that has a big enough context window for none of this to matter!