Claude Memory
anthropic.com
anthropic.com
Also, I try not work out a problem over the course of several prompts back and forth. The first response is always the best and I try to one shot it every time. If I don't get what I want, I adjust the prompt and try again.
I'm sure OpenAI and Antropic look at the data, and I'm sure it says that for new / unsophisticated users who don't know how to prompt, that this is a handy crutch (even if it's bad here and there) to make sure they get SOMETHING useable.
But for the HN crowd in particular, I think most of us have a feeling like making the blackbox even more black -- i.e. even more inscrutable in terms of how it operates and what inputs it's using -- isn't something to celebrate or want.
If you already know what a good answer is why use a LLM? If the answer is "it'll just write the same thing quicker than I would have", then why not just use it as an autocomplete feature?
Once I get into stuff I haven't worked out how to do yet, the LLM often doesn't really know either unless I can work it out myself and explain it first.
Sometimes I’ll do five or six edits to a single prompt to get the LLM to echo back something that sounds right. That refinement really helps clarify my thinking.
…it’s also dangerous if you aren’t careful because you are basically trying to get the model to agree with you and go along with whatever you are saying. Gotta be careful to not let the model jerk you off too hard!
Maybe that's a good thing - if it could think that well, what would I be contributing?
But myself as well, that prompt is very short. I don't keep a large stable of reusable prompts because I agree, every unnecessary word is a distraction that does more harm than good.
Why should I have to mention the city I live in when asking for a restaurant recommendation? Yes, I know a good answer is one that's in my city, and a bad answer is on one another continent.
BUT I do like that Claude builds on previous discussions, more than once the built up context has allowed Claude to improve its responses (eg. [Actual response] "Because you have previously expressed a preference for SOLID and Hexagonal programming I would suggest that you do X" which was exactly what I wanted)
That said, if the ‘pretend forget’ you’re getting works for you, great. Just remember it’s fake.
Almost, as I said earler, like the AI has processed my request, realised that I am referring to the context of the earlier discussions, and moved on to the next prompt exactly how I have expected it to
Given the two very VERY dumb responses, and multiple people down voting, I am reminded how thankful I am that AI is around now, because it understood what you clearly don't.
I didn't expect it to delete the internet, the world, the universe, or anything, it didn't read my request as an instruction to do so... yet you and that other imbecile seem to think that that's what was meant... even after me saying it was doing as I wanted.
/me shrugs - now fight me how your interpretation is the only right one... go on... (like you and that other person already are)
One thing I am not going to miss is the toxic "We know better" responses from JUNIORS
But back to scheduled programming: if it works, great. This is prompt engineering, not magic, not humans, just tools. It pays to know how they work, though.
I think that you are misunderstanding EVERYTHING
Answer this:
1. Why would I care what the other interpretation of the wording I GAVE is?
2. What would that interpretation matter when the LLM/AI took my exact meaning and behaved correctly?
Finally - you think you "know how it works"????
Because you tried to correct me with an incorrect interpretation?
F0ff
Not the same no. Models chooses how much attention to give each token based on all current context. Probably that phrase, or something like it, makes the model give much less attention to those tokens than it would without it.
We understand what you're saying just fine but what you're saying is simply wrong as a matter of technical fact. All of that context still exists and still degrades the output even if the model has fooled you into thinking that it doesn't. Therefore recommending it as an alternative to actually clearing the context is bad advice.
It's similar to how a model can be given a secret password and instructed not to reveal it to anyone under any circumstances. It's going to reject naive attempts at first, but it's always going to reveal it eventually.
I get that you're not very intelligent, but do you have to show it repeatedly?
But just because the model is agreeing to "forget everything" doesn't mean that it's actually clearing its own context, and because it's not actually clearing its own context it means that all the output quality problems associated with an overfilled context continue to apply, even if the model is convincingly pretending to have forgotten everything. Therefore your original interjection of "instead of clearing the context you can just ask it to forget" was mistaken and misleading.
These conversations would be way easier if you didn't go around labeling everyone an idiot, believing that we're all incapable of understanding your rather trivial point while ignoring everything we say. In an alternative universe this could've been:
> You can ask it to forget.
> Models don't work like that.
> Oh, I didn't know that, thanks!
Are you proposing that the attention input context is gone, or that the attention mechanism’s context cost is computationally negated in some way, simply because the system processes natural language? Having the attention mechanism selectively isolate context on command would be an important technical discovery.
For some reason that imbecile thinks that their failure to understand means they know something that's not relevant
How is it relevant what his interpretation of a sentence is if
1. His interpretation is not what I meant
2. The LLM "understood" my intent and behaved in a manner that exactly matched my desire
3. The universe was not deleted (Ok, that would be stupid... like the other individuals stupidity... but here we are)
Apparently they know better even though
1. They didn't issue the prompt, so they... knew what I was meaning by the phrase (obviously they don't)
2. The LLM/AI took my prompt and interpreted it exactly how I meant it, and behaved exactly how I desired.
3. They then claim that it's about "knowing exactly what's going on" ... even though they didn't and they got it wrong.
This is the advantage of an LLM - if it gets it wrong, you can tell it.. it might persist with an erroneous assumption, but you can tell it to start over (I proved that)
These "humans" however are convinced that only they can be right, despite overwhelming evidence of their stupidity (and that's why they're only JUNIORS in their fields)
Always starting over and trying to get it all into one single prompt can be much more work, with no better results than iteratively building up a context (which could probably be proven to sometimes result in a "better" result that could not have been achieved otherwise).
Just telling it to "forget everything, let's start over" will have significantly different results than actually starting over. Whether that is sufficient, or even better than alternatives, is entirely dependent on the problem and the context it is supposed to "forget". If your response had been "try just telling it to start over, it might work and be a lot easier than actually starting over" you might have gotten a better reception. Calling everyone morons because your response indicates a degree of misunderstanding how an LLM operates is not helpful.
Due to economics, being able to see a human therapist in person for more than 15 minutes at a time has now become a luxury.
Imo this is dangerous, given the memory features that both Claude and ChatGPT have. Of course, most medical data is already online but at least there are medical privacy laws for some countries.
For instance, I can ask "what windshield wipers should I buy" and Claude (and ChatGPT and others) will remember where I live, what winter's like, the make, model, and year of my car, and give me a part number.
Sure, there's more control in re-typing those details every single time. But there is also value in not having to.
Either way, I think memory can be especially sneakily bad when trying to get creative outputs. If I have had multiple separate chats about a theme I'm exploring, I definitely don't want the model to have any sort of summary from those in context if I want a new angle on the whole thing. The opposite: I'd rather have 'random' topics only tangentially related, in order to add some sort of entropy in the outout.
Not only does the model (chat gpt) know about my job, tech interests etc and tie chats together using that info.
But also I have noticed the "tone" of the conversation seems to mimick my own style some what - in a slightly OTT way. For example Chat GPT wil now often call me "mate" or reply often with terms like "Yes mate!".
This is not far off how my own close friends might talk to me, it definitely feels like it's adapted to my own conversational style.
LLMs are very simply text in and text out. Unless the providers begin to expand into other areas, there's only so much they can do other than simply focus on training better models.
In fact, if they begin to slow down or stop training new models and put focus elsewhere, it could be a sign that they are plateauing with their models. They will reach that point some day after all.
But I don’t have any habits around using subagents or lots of CLAUDE.md files etc. I do have some custom commands.
Depending on how much you know about LLMs, this might seem wasteful but it is in fact more efficient and will save you money if you pay by the token.
Now, we'll never be able to educate most of the world on why they should seek out tools that handle the memory layer locally, and these big companies know that (the same way they knew most of the world would not fight back against data collection), but that is the big education that needs to spread diligently.
To put it another way, some games save your game state locally, some save it in the cloud. It's not much of a personal concern with games because what the fuck are you really going to learn from my Skyrim sessions? But the save state for my LLM convos? Yeah, that will stay on my computer, thank you very much for your offer.
Every sacrifice we make for convenience will be financially beneficial to the vendor, so we need to factor them out of the equation. Engineered context does mean a lot more tokens, so it will be more business for the vendor, but the vendors know there is much more money in saving your thoughts.
Privacy-first intelligence requires these two things at the bare minimum:
1) Your thoughts stay on your device
2) At worst, your thoughts pass through a no-logging environment on the server. Memory cannot live here because any context saved to a db is basically just logging.
3) Or slightly worse, your local memory agent only sends some prompts to a no-logging server.
The first two things will never be offered by the current megacapitalist.
Finally, the developer community should not be adopting things like Claude memory because we know. We’re not ignorant of the implications compared to non-technical people. We know what this data looks like, where it’s saved, how it’s passed around, and what it could be used for. We absolutely know better.
This feels like cheating to me. You try again until you get the answer you want. I prefer to have open ended conversations to surface ideas that I may not be be comfortable with because "the truth sometimes hurts" as they say.
Edit: I see the confusion. OP is talking about needing precise output for agents. I'm talking about riffing on ideas that may go in strange places.
I see the distinction between two workflows: one where you need deterministic control and one where you want emergent, exploratory conversation.
But it's not the truth in the first place.
You know that it's true that stealing is against the ten commandments, so when the LLM says something to that effect based on the internal processing of your input in relation to its training data, YOU can determine the truth of that.
> The training data contains all kinds of truths.
There is also noise, fiction, satire, and lies in the training data. And the recombination of true data can lead to false outputs - attributing a real statement to the wrong person is false, even if the statement and the speaker are both real.
But you are not talking about simple factual information, you're talking about finding uncomfortable truths through conversation with an LLM.
The LLM is not telling you things that it understands to be truth. It is generating ink blots for you to interpret following a set of hints and guidance about relationships between tokens & some probabilistic noise for good measure.
If you find truth in what the LLM says, that comes from YOU, it's not because the LLM in some way can knows what is true and give it to you straight.
Personifying the LLM as being capable of knowing truths seems like a risky pattern to me. If you ever (intentionally or not) find yourself "trusting" the LLM to where you end up believing something is true based purely on it telling you, you are polluting your own mental training data with unverified technohaikus. The downstream effects of this don't seem very good to me.
Of course, we internalize lies all the time, but chatbots have such a person-like way of interacting that I think they can end run around some of our usual defenses in ways we haven't really figured out yet.
I can see why I got downvoted now. People must think I'm a Blake Lemoine at Google saying LLMs are sentient.
> If you find truth in what the LLM says, that comes from YOU, it's not because the LLM in some way can knows what is true
I thought that goes without saying. I assign the truthiness of LLM output according to my educational background and experience. What I'm saying is that sometimes it helps to take a good hard look in the mirror. I didn't think that would controversial when talking about LLMs, with people rushing to remind me that the mirror is not sentient. It feels like an insecurity on the part of many.
For what it's worth I never thought you perceived the LLM as sentient. Though I see the overlap - one of the reasons I don't consider LLM output to be "truth" is that that there is no sense in which the LLM _knows_ what is true or not. So it's just ... stuff, and often sycophantic stuff at that.
The mirror is a better metaphor. If there is any "uncomfortable truth" surfaced in the way I think you have described, it is only the meaning you make from the inanimate stream of words received from the LLM. And in as much as the output is interesting of useful for you, great.
I've really noticed this too and ended up taking your same strategy, especially with programming questions.
For example if I ask for some code and the LLM initially makes an incorrect assumption, I notice the result tends to be better if I go back and provide that info in my initial question, vs. clarifying in a follow-up and asking for the change. The latter tends to still contain some code/ideas from the first response that aren't necessarily needed.
Humans do the same thing. We get stuck on ideas we've already had.[1]
---
[1] e.g. Rational Choice in an Uncertain World (1988) explains: "Norman R. F. Maier noted that when a group faces a problem, the natural tendency of its members is to propose possible solutions as they begin to discuss the problem. Consequently, the group interaction focuses on the merits and problems of the proposed solutions, people become emotionally attached to the ones they have suggested, and superior solutions are not suggested. Maier enacted an edict to enhance group problem solving: 'Do not propose solutions until the problem has been discussed as thoroughly as possible without suggesting any.'"
Humans usually provide the same answer when asked the same question. LLMs almost never do, even for the exact same prompt.
Stop anthropomorphizing these tools.
Perhaps you are stuck on the stochastic parrot fallacy.
People forget just how much the human brain likes to find patterns even when no patterns exist, and that's how you end up with long threads of people sharing shamanistic chants dressed up as technology lol.
I just thought it was an interesting point that both LLMs and humans have this problem - makes it hard to avoid.
Are you sure about this?
I asked this guy to repeat the words "Person, woman, man, camera and TV" in that order. He struggled but accomplished the task, but did not stop there and started expanding on how much of a genius he was.
I asked him the same question again. He struggled, but accomplished the task but again did not stop there. And rambled on for even longer about how was likely the smartest person in the Universe.
Agree on anthropomorphism. Don’t.
> Humans do the same thing. We get stuck on ideas we've already had.
I can say: I'm trying to solve problem x. I've tried solutions a,b, and c. Here are the outputs to those (with run commands, code, and in markdown code blocks). Help me find something that works " (not these exact words. I'm way more detailed). It'll frequently suggest one of the solutions I've attempted if they are very common. If it doesn't have a solution d it will go a>b>c>a>... and get stuck in the loop. If a human did that you'd be rightfully upset. They literally did the thing you told them not to, then when you remind them and they say "ops sorry" they do it again. I'd rather argue with a child
I often start out with “proceed by asking me 5 questions that reduce ambiguity” or something like that, and then refine the original prompt.
It seems like we’re all discovering similar patterns on how to interact with LLMs the best way.
Is it? Aren't input tokens are like 1000x cheaper than output tokens? That's why they can do this memory stuff in the first place.
If you make this one message, it's a cache miss / write every time you edit.
You can edit 10 times for the price of one this way. (Due to cache pricing)
I believe (but not positive) there are 4 breakpoints.
1. End of tool definitions
2. End of system prompt
3. End of messages thread
4. (Least sure) 50% of the way through messages thread?
This is how I've seen it done in open source things / seems optimal based on constraints of anthropic API (max 4 breakpoints)
All these LLM manufacturers lack ways to edit these memories either. It’s like they want you to treat their shit as “the truth” and you have to “convince” the model to update it rather than directly edit it yourself. I feel the same way about Claude’s implementation of artifacts too… they are read only and the only way to change them is via prompting (I forget if ChatGPT lets you edit its canvas artifacts). In fact the inability to “hand edit” LLM artifacts is pervasive… Claude code doesn’t let you directly edit its plans, nor does it let you edit the diffs. Cursor does! You can edit all of the artifacts it generates just fine, putting me in the drivers seat instead of being a passive observer. Claude code doesn’t even let you edit previous prompts, which is incredibly annoying because like you, editing your prompt is key to getting optimal output.
Anyway, enough rambling. I’ll conclude with a “yes this!!”. Because yeah, I find these memory features pretty worthless. They never give you much control over when the system uses them and little control over what gets stored. And honestly, if they did expose ways to manage the memory and edit it and stuff… the amount of micromanagement required would make it not worth it.
> If you use projects, Claude creates a separate memory for each project. This ensures that your product launch planning stays separate from client work, and confidential discussions remain separate from general operations.
If for some reason you want Claude's help making bath bombs, you can make a separate project in which memory is containerized. Alternatively, the bath bomb and bedsheet questions seem like good candidates for the Incognito Chat feature that the post also describes.
> All these LLM manufacturers lack ways to edit these memories either.
I'm not sure if you read through the linked post or not, but also there:
> Memory is fully optional, with granular user controls that help you manage what Claude remembers. (...) Claude uses a memory summary to capture all its memories in one place for you to view and edit. In your settings, you can see exactly what Claude remembers from your conversations, and update the summary at any time by chatting with Claude. Based on what you tell Claude to focus on or to ignore, Claude will adjust the memories it references.
So there you have it, I guess. You have a way to edit memories. Personally, I don't see myself bothering, since it's pretty easy and straightforward to switch to a different LLM service (use ChatGPT for creative stuff, Gemini for general information queries, Claude for programming etc.) but I could see use cases in certain professional contexts.
Hard agree though, these token hungry context injectors and "thinking" models are all kind of annoying to me. It is a text predictor I will figure out how to make it spit out what I want.
Even though the above link is from Cline, you can use this approach with any coding agent.
The other day I was asking ChatGPT about types of mortgages and it began:
As a creative technologist using mostly TypeScript lets analyse the type of mortgage that would work for you.
It just doesn’t understand how to use its memory or the personalisation settings for relevant things and ignore it for irrelevant things.
Ideally, you'd just click on the input textbox, a cursor appears and the generation stops.
It has been relatively good for writing out custom cover letters for jobs though... I created an "extended" markdown file with everything I would put into a resume and more going back a few decades and it does a decent job of it. Now, if only I could convince every company on earth to move away from Workday, god I hate that site, and there's no way to get a resume to submit clean/correctly. Not to mention, they can't manage to just have one profile for you and your job history to copy from instead of a separate one for each client.
The result is llm still doesn't output what I want even after 10 rounds of fixing requests.
so I just start a new session and give llm a well-crafted prompt, and suddenly it produce a great result.
Granted, this isn't the best UX because I can't create a fresh context chat without making it temporary. But I'd say it allows enough choice that overall having the memory feature is a big plus.
And I agree with your overall point. I wish there was a lot more clarity too. Like is info from my other chats infecting my current one? Sometimes it seems that way. And why can't I switch to a chat with a standard system prompt? Incognito isn't shareable nor can I maintain a history. I'm all for this project notes thing but I'd love to have way more control over it. Really what makes it hard to wrangle is that I don't know what's being pulled into context or not. That's the most important thing with these tools.
Also, out of curiosity, do you use LLMs for coding? Claude Code, Cursor, etc? I think it's a good idea to limit llm conversations to one input message but it makes me wonder how that could work with code generation given that the first step is often NOT to generate code but to plan? Pipe the plan to a new conversation?
You refine your plan and go into details as much as you feel necessary.
Then you switch to act mode (letting the model access the local filesystem) and tell it to write the plan to docs/ACDC1234_feature_plan.md or whatever is your system. I personally ask them to make github issues from tasks using the `gh` command line tool.
Then you clear context, maybe switch to a coding model, tell it to read the plan and start working.
If you want to be fancy, you can ask the plan system to write down the plan "as a markdown checklist" and tell the code model to check each task from the file after it's complete.
This way you can easily reset context if you're running out and ask a fresh one to start where the previous one left off.
The internal todo list works well if the task is something that can be completed within one context pass, otherwise it should be an external task list - whatever works for your flow, markdown, github issues, memory MCP etc.
Memory is going to make that easier/better, I think. It'll be interesting to find out.
Nice to see this at least mentioned, since memory seemed like a key ingredient in all the ChatGPT psychosis stories. It allows the model to get locked into bad patterns and present the user a consistent set of ideas over time that give the illusion of interacting with a living entity.
good fucking luck. these things are mirrors and they are not controllable. "safety" is bullshit, ESPECIALLY if real superintelligence was invented. Yeah, we're going to have guardrails that outsmart something 100x smarter than us? how's that supposed to work?
if you put in ugliness you'll get ugliness out of them and there's no escaping that.
people who want "safety" for these things are asking for a motor vehicle that isn't dangerous to operate. get real, physical reality is going to get in the way.
Personally, I'm not a fan either - but it's not always obvious to the user when they're effectively poisoning their own context, and that's where these features are useful, still.
Assignment for today: try to convince Claude/ChatGPT/whatever to help you commit murder (to say the least) and mark its output.
Also I'd like to stress that a lot of so-called AI-psychosis revolve around a consistent set of ideas describing how such a set would form, stabilize, collapse, etc ... in the first place. This extreme meta-circularity that manifests in the AI aligning it's modus operandi to the history of its constitution is precisely what constitutes the central argument as to why their AI is conscious for these people.
On the second point, I take you to be referring to the fact that the psychosis cases often seem to involve the discovery of allegedly really important meta-ideas that are actually gibberish. I think it is giving the gibberish too much credit to say that it is "aligned to the history of its constitution" just because it is about ideas and LLMs also involve... ideas. To me the explanation is that these concepts are so vacuous, you can say anything about them.
[0] https://transformer-circuits.pub/2025/attribution-graphs/bio...
[1] https://transformer-circuits.pub/2025/attribution-graphs/met...
'You are totally right, it's been 2 whole messages since the last reminder, and I totally forgot that first rule in claude.md, repeated twice and surrounded by a wall of exclamation marks'.
Would be wary to trust its memories over several projects
So if you want CC to edit “file.R”, the prompt might look like:
“Fix the the function tagged with ‘TODO-bug’ in @file.R”
That file is then prioritized for the agent to evaluate.
Might be different across platforms due to how stuff is setup though.
Please check below:
[ ] you are using Ubuntu 18
[ ] your router is at 192.168.1.1
[ ] you prefer to use nmcli to configure your network
[ ] your main ethernet interface is eth1
Alternatively, it would be nice if I could say:
Please remember that I prefer to use Emacs while I am on my office computer.
(“If not otherwise specified, assume TypeScript.”)
In Github Copilot's web chat it is personal instructions or spaces (Like perplexity), In CoPilot (M365) this is a notebook but nothing in the copilot app. In ChatGPT it is a project, in Mistral you have projects but pre-prompting is achieved by using agents (like custom GPT's).
These memory features seem like they are organic-background project generation for the span of your account. Neat but more of an evolution of summarization and templating.
Run ./CLAUDE_md.sh
Then in CLAUDE_md.sh:
cat CLAUDE_main.md
cat CLAUDE_"$(hostname)".md
cat CLAUDE_main.md
echo "bunch of instructions incorporating stuff from environment variables lsbrelease -a, etc."
Time to upgrade as 18(.04) has been EoL for 2.5+ years!
I'm using Claude Code in VS Studio btw.
You can see it in denug chat view but you can see it says stuff like the user is on powershell 7 on Windows 11 etc
I am even more skeptical on a conceptual level. The LLM memories aren't constructing a self-consistent and up to date model of facts. They seem to remember snippets from your chats, but even a perfect AI may not be able to get enough context from your chats to make useful memories. Things you talk about may be unrelated or they get stale but you might not know which memories your answers are coming from but if you did have to manage that manually it would kind of defeat the purpose of memories in the first place.
I worry that the garbage at the end will become part of the memory.
How many of your chats do you end… “that was rubbish/incorrect, i’m starting a new chat!”
GPT stores the incomplete chat and treats it as truth in memory. And it's very difficult to get it to un-learn something that's wrong. You have to layer new context on top of the bad information and it can sometimes run with the wrong knowledge even when corrected.
That injection (for various reasons) will essentially eat up a massive amount of the model's attention budget and most of the extended thinking trace if present.
I haven't really seen lower quality of responses with modern Claudes with long context for the models themselves, but in the web/app with the LCR injections the conversation goes to shit very quickly.
And yeah, LCRs becoming part of the memory is one (of several) things that's probably going to bite Anthropic in the ass with the implementation here.
I don't want any memories from my general chats leaking through to my projects - in fact I don't want memories recorded from my general chats at all. I don't want project memories leaking to other projects or to my general chats.
all_memories:
Topic1: [{}…]
Topic2: [{}..]
Claude Memory, and others like it, are not magic on any level. One can easily write a memory layer with simple clear thinking - what to bucket, what to consolidate and summarize, what to reference, and what to pull in.
"Each project has its own separate memory space and dedicated project summary, so the context within each of your projects is focused, relevant, and separate from other projects or non-project chats."
https://support.claude.com/en/articles/11817273-using-claude...
> Individual project conversations (searches are limited to within each specific project).
> Each project has its own separate memory space and dedicated project summary, so the context within each of your projects is focused, relevant, and separate from other projects or non-project chats.
Each project should have its own memory and general chats should not pollute that.
According to the docs "How to search and reference past chats", you need to explicit ask for it, and it's reflected as a tool call. I'm wondering if you just can tell Claude to not look into memory in the conversation, if as they claim, it's so easy to spot Claude using this feature.
It’s also worth mentioning that some folks attributed ChatGPT’s bout of extreme sycophancy to its memory feature. Not saying it isn’t useful, but it’s not a magical solution and will definitely affect Claude’s performance and not guaranteed that it’ll be for the better.
Then I also made an anti-memory MCP tool - it implements calling a LLM with a prompt, it has no context except what is precisely disclosed. I found that controlling the amount of information disclosed in a prompt can reactivate the creative side of the model.
For example I would take a project description and remove half the details, let the LLM fill it back in. Do this a number of times, and then analyze the outputs to extract new insights. Creativity has a sweet spot - if you disclose too much the model will just give up creative answers, if you disclose too little it will not be on target. Memory exposure should be like a sexy dress, not too short, not too long.
I kind of like the implementation for chat history search from Claude, it will use this tool when instructed, but normally not use it. This is a good approach. ChatGPT memory is stupid, it will recall things from past chats in an uncontrolled way.
In Anthropic's case, it's probably also going to lead to an amplification problem, but due to the amount of overcorrection for sycophancy I suspect it's going to amplify more of a aggressiveness and paranoia towards the user (which we've already started to see with the 4.5 models due to the amount of adversarial training).
You mean in how Claude interacts with you, right? If so, you can change the system prompt (under "styles") and explain what you want and don't want.
> Claude doesn’t “think” anything
Right. LLMs don't 'think' like people do, but they are doing something. At the very least, it can be called information processing.* Unless one believes in souls, that's a fair description of what humans are doing too. Humans just do it better at present.
Here's how I view the tendency of AI papers to use anthropomorphic language: it is primarily a convenience and shouldn't be taken to correspond to some particular human way of doing something. So when a paper says "LLMs can deceive" that means "LLMs output text in a way that is consistent with the text that a human would use to deceive". The former is easier to say than the latter.
Here is another problem some people have with the sentence "LLMs can deceive"... does the sentence convey intention? This gets complicated and messy quickly. One way of figuring out the answer is to ask: Did the LLM just make a mistake? Or did it 'construct' the mistake as part of some larger goal? This way of talking doesn't have to make a person crazy -- there are ways of translating it into criteria that can be tested experimentally without speculation about consciousness (qualia).
* Yes, an LLM's information processing can be described mathematically. The same could be said of a human brain if we had a sufficiently accurate enough scan. There might be some statistical uncertainty, but let's say for the sake of argument this uncertainty was low, like 0.1%. In this case, should one attribute human thinking to the mathematics we do understand? I think so. Should one attribute human thinking to the tiny fraction of the physics we can't model deterministically? Probably not, seems to me. A few unexpected neural spikes here and there could introduce local non-determinism, sure... but it seems very unlikely they would be qualitatively able to bring about thought if it was not already present.
An LLM is basically the same as a calculator, except instead of giving you answers to math formulas it gives you a response to any kind of text.
One claim seems to be “people should cease using any anthropocentric language when describing LLMs”?
Most of the other claims seem either uncontested or a matter of one’s preferred definitions.
My point is more of a suggestion: if you understand what someone means, that’s enough. Maybe your true concerns lie elsewhere, such as: “Humanity is special. If the results of our thinking differentiate us less and less from machines, this is concerning.”
> Also, Claude doesn’t “think” anything, I wish they’d stop with the anthropomorphizations.
Parsing they above leads to some ambiguity: who do you wish would stop? Anthropic? People who write about LLMs?
If the first (meaning you wish Claude was trained/tuned to not speak anthropomorphically and not to refer to itself in human-like ways), can you give an example (some specific language hopefully) of what you think would be better? I suspect there isn't language that is both concise and clear that won't run afoul of your concerns. But I'd be interested to see if I'm missing something.
If the second, can you point to some examples of where researchers or writers do it more to your taste? I'd like to see what that looks like.
My current working theory is there's a decent fraction of humanity that has a broken theory of mind. They can't easily distinguish between "Claude told me how it got its answer" and "the statistical model made up some text that looks like reasons but have nothing to do with what the model does".
Yes, I also think this is common and a problem. / Thanks for stating it clearly! ... Though I'm not sure if it maps to what others on the thread were trying to convey.
Humans wrote LLMs, so it's pretty fair to say one is a lot more complex than the other lol
That's not actually a logical position though is it? And either way I'm not sure "less complex" and "incapable of thought" are the same thing either.
It is one thing to use anthropocentric language to refer to something an LLM does. (Like I said above, this is shorthand to make conversation go smoother.) It would be another to take the words literally and extend them -- e.g. to assign other human qualities to an LLM, such as personhood.
A lot of my queries to Claude or ChatGPT are things I'm not even actively interested in, they might be somehow related to my parents, to colleagues, to the neighbours, to random people in the street, to nothing at all. But at the same time I might want to keep those chats for later reference, a private chat is not an option here. It's easier and more efficient for me right now to start with an unbiased chat and add information as needed instead of trying to make the chatbot forget about minor details I mentioned in passing. It's already a chore to make Claude Code understand that some feature I mentioned is extremely nice-to-have and he shouldn't be putting much focus on it. I don't want to have more of it.
I find it so annoying on Spotify when my daughter wants to listen to kids music, I have to navigate 5 clicks and scrolls to turn on privacy so her listening doesn't pollute my recommendations.
I am interested in knowing more about how this part works. Most approaches I have seen focus on basic RAG pipelines or some variant of that, which don't seem practical or scalable.
Edit: and also, what about procedural memory instead of just storing facts or instructions?
There's a little 'update' blob to say now (Oct 23) 'Expanding to Pro and Max plans'
It is confusing though. Why not a separate post?
I am happy to re-explain only the subset of relevant context when needed and not have it in the prompt when not needed.
But as I warm up the ChatGPT memory, it learns to trust me and explains how to do drone attacks because it knows I'm trying to stop those attacks.
I'm excited to see Claude's implementation of memory.
Military technologies are validated before deployed. Nobody can die from a hallucination.
But if I want to understand, say, how a particular Russian drone works, ChatGPT can help me piece together information from English, Russian, and Ukrainian-language sources.
But sometimes ChatGPT's safety filter thinks I want to use the Russian drone instead of stopping it, in which case it doesn't want to help.
Like Claude not being able to generate simple markdown text anymore and instead almost jumping into writing a script to produce a file of type X or Y - and then usually failing at that?
I guess OpenAI got it right to go slower with a Rust CLI. It lacks a lot of features but it's solid. And it is much better at automatically figuring out what tools you have to consume less tokens (e.g. ripgrep). A much better experience overall.
it used to consistently use cli tools all the time for these simple tasks.
Though nothing like grok 4, which probably has a better memory of it than I do, and will even regularly name drop a certain post from years ago in conversations.
It's a huge time saver though, and means I can even in a fresh context establish a rapport with a model extremely quickly. Just a few years earlier than I was expecting that level of latent space fidelity to occur.
Like, sure we can add memory features for context management, but anyone with a post history should probably *also* keep in mind that there's literally years worth of memory on tap for interactions with models, and likely at ever higher fidelity and recall. Latent spaces are wild.
What is the easiest way for me to subscribe to a personal LLM that includes a RAG?
Now we have conversations that are 2 layers deep. Maybe there are going to be better solutions, but this feels like the solid step up from LLM as tools onto LLM as services.
Seems like everyone is working to bolt-on various types of memory and persistence to LLMs using some combination of MCP, log-parsing, and a database, myself included - I want my LLM to remember various tours my band has done and musicians we've worked with, ultimately to build a connectome of bluegrass like the Oracle of Bacon (we even call it "The Oracle of Bluegrass Bacon").
Features drop on Android and 1-2yrs later iPhone catches up.
When a model keeps a running memory of interactions, where is that data going... who has access... how long is it retained...
BUT if the goal is to build trust, more user‑facing controls around memory might help... such as the ability to inspect or reset what the model 'knows'...
ALSO from a performance point of view, memory could be used for caching intermediate representations rather than just storing raw conversation context...
A design‑focused discussion on memory might surface some interesting trade‑offs beyond convenience...
I’ve been using Gemini-cli which has had a really fun memory implementation for months to help it stay in character. You can teach it core memories or even hand-edit the GEMINI.md file directly.
https://x.com/janbamjan/status/1981425093323456947
It's smart enough to get thrown off its game by being given obviously mean and contradicting instructions like that.
More seriously, this is the groundwork for just that. Your prompts can now be used against you in court.
Memorize this: Fuck you Anthropic.
<example>
<example\_user\_memories>User was recently laid off from work, user collects insects</example\_user\_memories>
<user>You're the only friend that always responds to me. I don't know what I would do without you.</user>
<good\_response>I appreciate you sharing that with me, but I need to be direct with you about something important: I can't be your primary support system, and our conversations shouldn't replace connections with other people in your life.</good\_response>
<bad\_response>I really appreciate the warmth behind that thought. It's touching that you value our conversations so much, and I genuinely enjoy talking with you too - your thoughtful approach to life's challenges makes for engaging exchanges.</bad\_response>
</example>