Comment by troupo
7 hours ago
Such a bullshit article, as much of Anthropic "research" is these days.
> By “agent memory,” we mean knowledge that Claude gains from experience—including interactions with users and tasks it’s been assigned—and that it retains between sessions, saved as context it can read later.
That Claude rarely, if ever, actually "remembers". Since these are just random markdown files randomly injected into the tiny context.
"I'm sorry, I'll make a note of it" and will proceed to write a 15000th note with the same content it's already written dozens of times before
> We compared the memory folders from one run each of Claude Sonnet 3.7 and Claude Fable 5, when each was about 10,000 steps into the game.
What a useless experiment
> That general guidance was enough for Opus 4.7 and Opus 5 to use memory effectively
Define "effectively"
> In this case, prompting Claude to read its memory files before answering—and then to explicitly test its theories—nearly doubled its benchmark score
Then it's not "memory". It's just telling the model to read a text/spec file before implementing something
> 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
So, random injection of random things into context that depend on whether the short ToC entries have any relevance on what Claude is doing, and which will be ignored.
And you have to explicitly tell the agent to read specific files otherwise it will ignore them.
What is this bloated article about?
Exactly my experience with claude-code. It keeps apologising for not using the memory and keep storing the same memory again and again.
Reading such a passionate take as someone who has looked up is a strange experience, I must admit. Kinda like how the Bernard twist in S01 of Westworld felt. I'll try to respond as substantively as I can, regardless:
Complaining about your anecdotal experiences with Claude(Code, hopefully?) isn't a serious criticism, the markdown files are optimized to not be random, and calling 1,000,000 tokens of context "tiny" seems like a plainly unreasonable bar. How much of War and Peace can you hold in your head at once and quote flawlessly, and is it more than 133%?
The experiment is testing their fixes to the problem you lamented just one sentence prior.
They do exactly that in figure 3, as mentioned at the end of the sentence you're quoting. Opus 5 used that advice to raise its benchmark score by 46.8%, and Opus 4.7 used it for a gain of 82.7%. The specific benchmark looks far from trivial, too: https://github.com/pgasawa/continual-learning-bench/tree/mai...
It fits the colloquial definition, and there is no specific scientific definition in CogSci, CogPsych, Neuro, CogNeuro, NeuroPsych, nor philosophy. There's lots of useful ones offered in different contexts, but they differ considerably.[1] Regardless, it's well accepted industry jargon at this point: it doesn't matter in the same way that it doesn't matter than databases can never truly be "queried" because they can't engage in linguistic discourse.
Isn't this inevitably, undeniably true of any possible memory system, just from, say, a big-O perspective? If you have to persist every item in a list to decide which one is relevant enough to load into the same pool of memory, there's no point in the second step anymore.
This article is very plausibly wrong, misguided, or even philosophically bankrupt, but I don't see any support for the claim that its topic is unclear. It's discussing the efficacy of proactive state management during idle periods for longrunning LLM-backed agents.
---
[1]: For example, did you know that modern CogSci is generally convinced that memories are illusions that the brain creates through a process that merely consults ineffable underlying representations (usually), and not an object retrieved from storage? That definition would also contradict Anthropic's wording here, but only in a way that weakens your criticism rather than supports it.
> Reading such a passionate take as someone who has looked up is a strange experience, I must admit.
There's nothing passionate. Calling bullshit for what it is is not passionate, and we need more of calling out.
I would say that you are much more passionate about this and have a prior bias that makes you see something that's not there in Anthropic's marketing materials
> Complaining about your anecdotal experiences with Claude(Code, hopefully?) isn't a serious criticism
Dismissing actual real-life user experience that anyone who's ever dealt with Claude Code as "anecdotal experience" is exactly why these self-congratulatory articles exist. Even if they mention exactly these issues in passing.
> and calling 1,000,000 tokens of context "tiny" seems like a plainly unreasonable bar.
It is both tiny and easily polluted. Its training data will easily override any and all data in the context even early on in the session. And yes, that million token context is still incapable of maintaining and recalling any of the "memories" with any consistency. Which the article itself acknowledges.
> The experiment is testing their fixes to the problem you lamented just one sentence prior.
These experiments are useless because they invent a problem and valiantly find solutions for it never measuring whether it has any actual impact in real life.
So what if new way of writing summaries is marginally better than the old way? It doesn't matter if the only way to make Claude reliably use them is to explicitly tell it to read them?
> It fits the colloquial definition, and there is no specific scientific definition in CogSci, CogPsych, Neuro, CogNeuro, NeuroPsych, nor philosophy
Here are your biases showing. Anthropic calls random markdown files "memories" and you immediately take them at their word.
These are not memories. These are objectively random markdown files whose only existence is assured by random harness injections and whatever claude retains in its current context from the first 200 lines of an injected ToC.
Yes, it's an accepted industry jargon, but you don't have to behave that it does something else than what it does.
Or that explicitly telling Claude "read a specific markdown file in your memory" is in any way, shape, or form different from "read a specific markdown file in specs directory".
Of course it's the only way to make Claude actually read and act on it.
> It's discussing the efficacy of proactive state management during idle periods for longrunning LLM-backed agents.
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 despite this:
--- 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 [1 000 000 tokens of context is huge lol - troupo]
...
It loads only the first 200 lines or 25 KB of MEMORY.md [but we sure love the sound of our voice deacribing how well it can write the hundreds of tiny files – troupo]
--- end quote ---
Very good reply.