  There's nothing passionate. Calling bullshit for what it is is not passionate, and we need more of calling out.

...you reacted to this blogpost by calling it "bullshit" and "bloated", and criticizing it with obvious hyperbole ("random markdown files randomly injectect into the tiny context", "15000th note", "dozens of times before", "useless experiment", "random injection").

You're allowed to feel emotion -- that wasn't a refutation. But denying that you indeed are typing with responding seems plainly obstinate, sorry. Again I will do my best to respond in line with the rules of this forum, even if I don't feel the same grace being given in return.

  I would say that you are much more passionate

What makes you say that? Because I disagree? A comment starting with "I'll try to reply as substantively as I can" isn't exactly firey, much less one that makes pain to cede ground and offer friendly offramps to the polemics.

  Dismissing actual real-life user experience

Well I have positive actual real-life user experience with these products. It seems we're at an impasse, there! 1-1=0, after all. This is why anecdotal evidence isn't valid, generally speaking.

  It is both tiny and easily polluted

This just comes down to "nuh-uh". So: yeah-huh! 133% of War and Peace is a lot of text to hold at the forefront of one's mind simultaneously.

   that million token context is still incapable of maintaining and recalling any of the "memories" with any consistency. Which the article itself acknowledges.

That's the exact opposite of the findings of the experiment that you yourself quoted: achieving an 82% improvement on a challenging benchmark without detailed instructions for memory use is in no sense "incapable".

  Here are your biases showing. Anthropic calls random markdown files "memories" and you immediately take them at their word.

You're not taking the time to read what I'm writing :( "It fits the colloquial definition" objectively isn't taking Anthropic "at their word". On an absolute, undeniable, factual level, this attempted ad-hominem is based on a faulty premise.

  These are not memories. These are objectively random markdown files

What are memories, and why does text stored in a computer "objectively" not fit that definition? Even if you offered one now (which, again, would be scientifically unsound), it'd be too late; it's beyond obvious that this reasoning is a-posteriori.

  Yes, it's an accepted industry jargon, but you don't have to behave that it does something else than what it does.

How am I doing that, specifically? You're the one taking issue with the industry jargon, and I offered no description of "what it does" in the first place, much less one that conflicts with some singular objective definition of "memory" (which, again, does not exist).

  Of course it's the only way to make Claude actually read and act on it.

This is objectively incorrect on an empirical level, as proven in the article we are commenting under.

  No. No it doesn't. It considers "components of agent memory" and focuses solely on "how we can write markdown files", and spends probably less than 1% on "how we can make sure the agent actually uses them". Even the conclusion completely ignores it

A) The desciption I gave does not require this arbitrary focus of yours.

B) Again, every single empirical result in the entire paper relies on the mechanism that you find to be non-existent.

▲ troupo 8 hours ago | parent [-]

> Well I have positive actual real-life user experience with these products.

I never said I didn't have a positive real-life user experience with these products.

> So: yeah-huh! 133% of War and Peace is a lot of text to hold at the forefront of one's mind simultaneously.

Let me quote the article again:

--- start quote ---

Injected memories use up valuable tokens, and an incorrect note can be carried into every session. Loading a note also doesn’t guarantee that Claude acts on it: our documentation on Claude Code warns that longer instruction files reduce adherence

--- end quote ---

Both the article and the linked docs say that Claude will only load ~200 lines of memory index (not the entire memory), and that you should keep your Claude.md below 200 lines.

Never knew War and Peace was only 400 lines long.

> What are memories, and why does text stored in a computer "objectively" not fit that definition?

Memories is something a model wouldn't need to re-read from scratch every time. Memories isn't something you would need to be always told to reliably use.

And no, whatever human analog you come up with isn't relevant, as we're not talking about humans.

> This is objectively incorrect on an empirical level, as proven in the article we are commenting under.

Let me quote the article:

--- start quote ---

Loading a note also doesn’t guarantee that Claude acts on it: our documentation on Claude Code warns that longer instruction files reduce adherence.

It loads only the first 200 lines or 25 KB of MEMORY.md

--- end quote ---

> Again, every single empirical result in the entire paper relies on the mechanism that you find to be non-existent.

1. I never said it was inexistent

2. The "empirical results" in the paper completely ignore the actual limitations of "memory". It explicitly says that "nope, Claude will not use memory because of known limitations", and immediately sweeps past it as if it wasn't important.

The entirety of "dreaming" and "memory" is completely negated by this:

--- start quote ---

Loading a note also doesn’t guarantee that Claude acts on it: our documentation on Claude Code warns that longer instruction files reduce adherence.

Claude doesn’t load all of its auto memory into every session’s context window. It loads only the first 200 lines or 25 KB of MEMORY.md

--- end quote ---

And this can be literally confirmed empirically by anyone actually working with these tools and watching what they do. E.g., Claude will completely ignore most "memories" and write the same "memory" again and again and again in potentially dozens of files when you tell it that something is wrong. Often repeating what is already there in explicit CLAUDE.md instuctions.

Oh wait. The amazing 1 000 000 token context that can seemingly hold all of "War and Peace" cannot reliably hold more than 200 lines of explicit instructions and more than 200 lines of "memory".

▲ bbor 6 hours ago | parent [-]

  I never said I didn't have a positive real-life user experience with these products.

Irrelevant; you were the one who was raising anecdotal evidence and defending it when challenged. Also c'mon surely you realize that you're just disagreeing for the sake of it by this point?

  Both the article and the linked docs say that Claude will only load ~200 lines of memory index

Irrelevant; you were discussing the context window itself, not how many lines of text are loaded per note.

  Memories isn't something you would need to be always told to reliably use.

A) Not a definition, and, again, B) Disproven directly by the findings I jsut quoted to you.

  And no, whatever human analog you come up with isn't relevant, as we're not talking about humans.

That's how analogues work.

  Let me quote the article:

Irrelevant quote.

   I never said it was inexistent

...again, surely even you realize that quibbling over 1% present and 0% present isn't productive, in this context? What's the point?

  It explicitly says that "nope, Claude will not use memory because of known limitations"

The entire article is about Claude using memory.

  Claude will completely ignore most "memories" and write the same "memory" again and again and again in potentially dozens of files when you tell it that something is wrong.

Is it possible that you haven't used Claude since Sonnet 3.7? Cause, again, this anecdotal experience isn't reflective of today's reality. And why would anyone care about the anecdotal experience of someone who defines themselves by hating LLMs in the first place?

  The amazing 1 000 000 token context that can seemingly hold all of "War and Peace" cannot reliably hold more than 200 lines of explicit instructions and more than 200 lines of "memory".

All of War and Peace could fit into one line. Regardless, 1M is 1M is 1M. This isn't a thing you can have a take about, it's a plain fact. You're able to feed many thousands of sentences of instructions and memories if you'd like to, it's just inefficient.

	▲	troupo 5 hours ago \| parent [-]
		> Irrelevant; you were the one who was raising anecdotal evidence and defending it when challenged. That "anecdotal" evidence is literally explicitly stated in the article > Irrelevant; you were discussing the context window itself, not how many lines of text are loaded per note. Literally quoting the article: "longer instruction files reduce adherence" It doesn't matter if the context is 1mln tokens if a file with more than 200 lines can lead to loss of adherence and context pollution. It doesn't matter if it is marketed as having the ability to hold the entirety of War and Piece if in reality it cannot hold much without being polluted. > That's how analogues work. Analogues never work. > Irrelevant quote. That is basically the entirety of your entire "argument". "What I say is correct, and Anthropic's article is seminal. Whatever you say, including quoting Anthropic's own materials is irrelevant if it doesn't fit my pre-biased worldview." > The entire article is about Claude using memory. No, it's not really about that once you cut to the meat and heart of it. > Is it possible that you haven't used Claude since Sonnet 3.7? Is is possible for you to not assign your opponents thoughts and words while completely dismissing what they are actually telling you? Apparently not. > You're able to feed many thousands of sentences of instructions and memories if you'd like to, it's just inefficient. Lol. So the size of the context is irrelevant when it's inneficient, and the models apparently lose adherence at as much as 200 lines of instructions. But sure, calling this "tiny context" somehow insults your sensibilities. Anyway. I've seen enough of your "arguments" to know I couldn't care less to continue this conversation. Adieu.