Could someone explain the appeal of account-wide memory to me? Anthropic’s marketing indicates that nothing bleeds over, but I’m just so protective of my context that I cannot imagine having even a majorly distilled version of my other chats and preferences having on weight on the output. As for certain preferences like code styling or response length, these are all fit for custom instructions, with more detailed things in Skills. Ultimately like many things in LLM web UX, it seems to cater to how the masses use these tools.
I use Claude code in a number of different parts of my business - coding internal applications, acting as a direct interface to SaaS via APIs and just general internal use.
I find there is a virtuous cycle here where the more I use it, the more helpful it is. I fired my bookkeeper and have been using Claude with a QBO API key instead, and because it already had that context (along with other related business context), when I gave it the tax docs I gave to my CPA for 2024's taxes plus my return, and asked it to find mistakes, it determined that he did not depreciate goodwill from an acquisition. CPA confirmed this was his error and is amending my return.
Then I thought it'd be fun to see how it would do on constructing my 2024 return just from the same source docs my CPA had. First time I did it, it worked for an hour then said it had generated the return, checked it against the 2024 numbers and found they're the same. I had removed the 2024 before having it do this to avoid poisoning the context with the answers, but it turns out it had a worksheet .md file that it was using on prior questions that I had not erased (and then it admitted that it had started from the correct numbers).
In order to make sure I wouldn't have that issue again, I tried the 2024 return again, completely devoid of any historical context in a folder totally outside of my usual Claude Code folder tree. It actually got my return almost entirely correct, but it missed the very same deduction that it had caught my CPA missing earlier.
So for me, the buildup of context over time is fantastic and really leads to better results.
I already switched to claude a while ago. Didn’t bring along any context, just switched subscriptions, walked away from chatgpt and haven’t touched it again. Turned out to be a non-event, there really is no moat.
I switched not because I thought Claude was better at doing the things I want. I switched because I have come to believe OpenAI are a bad actor and I do not want to support them in any way. I’m pretty sure they would allow AGI to be used for truly evil purposes, and the events of this week have only convinced me further.
If Claude could stay available I might consider it. Unfortunately right now, out of the big three, only Gemini has reliable uptime. As much as I dislike Google it's the only reliable option.
I'm very curious, will OpenAI basically block "I'm moving to another service and need to export my data. List every memory you have stored about me, ..." and similar, if so how and why?
It's very interesting to learn more about because it challenges 1 core aspect of the economical competition : the moat.
If one can literally swap one AI service for another, then where does the valuation (and the power that comes with it) come from?
PS: I'm not interested in the service itself as I believe the side effects of large scale for-profit are too serious (and I don't mean doomdays AI takeover, I simply mean abuse of power, working conditions, downskilling, political influence as current contracts with US defense are being made, ads, ecological, etc) to be ignored.
On a related note, I have been experimenting with a small prototype for cross-agent, device-local active memory called brAIn (https://github.com/glthr/brAIn). It delivers a personalized agent experience with everything stored locally in a single file (agent.brain), and supports reusing semantic memory across projects. In practice, this means brAIn can identify and apply behavioral patterns you have used in other contexts whenever they are relevant. (I realize the repository should include a concrete example of this, and I will update it today to add one).
Memory in general Chat apps is actually more harmful than helpful imo.
It biases the LLM responses to your background which has the same effect as filter bubbles. You end up getting your own thoughts spit back at you.
Of course sometimes this is useful if you only use your chatbot to ask personal things like: "What should I eat today?".
But if you use it for anything else you're much better off having full control over the prompt. I can always say: "Hey btw I am german and heavily anti surveillance, what should I know about the recent anthropic DoW situation?" but with memory I lose the option of leaving out that first part.
I tried all of Codex, OpenCode, Claude Code and Cursor these past few weeks. It was surprising to me that all of them have slightly different conventions for where to put skills, how to format MCP servers (how environment variables need to be specified etc), what the AGENTS/CLAUDE file needs to be called, what plugins/marketplaces are...it's a big mess for anyone trying to have a portable config in their dotfiles that can universally apply to any current and future agent.
It also showed me the difference between expectation and reality...even though these are billion dollar companies, they still haven't figured out how to make lag-free TUIs, non-Electron apps, or even respect XDG_CONFIG. The focus is definitely more on speed and stuffing these tools full of new discoveries and features right now
There's a bit of psychology around models vs. harnesses as well. You can't shake off the feeling that maybe Claude would perform better in its native harness compared to VSCode/OpenCode. Especially because they've got so many hidden skills (like the recently introduced /batch), that seem baked into the binary?
The last thing I can't figure out is computer use. Apparently all the vendors say that their models can use a mouse and keyboard, but outside of the agent-browser skill (which presumably uses playwright), I can't figure out what the special sauce is that the Cloud versions of these Agents are using to exercise programs in a VM. That is another reason why there is a switching cost between vendors.
Being able to import context and preferences from other AI providers in one step saves a lot of time, especially for ongoing projects. It makes Claude feel seamless and continuity-friendly. Having this on all paid plans adds great value for heavy users.
This method of copying an LLM-generated summary of your preferences into Claude memory feels similar to their recommendation to use /init to generate a CLAUDE.md based on the project, which recent research[0] suggests may be counterproductive.
I would assume both Claude memory and CLAUDE.md work best when they're carefully curated, only containing what you've found yourself having to repeat.
At least as an EU user I was also able to export ALL my data, audio files images etc in one zip. Took exactly (on the minute) 24 hours for the download link to arrive but hey.
This way you can have Claude distill the memory as you wish.
I got very excited when I saw this title, because I've wanted to consolidate on Claude for a long time. I have been using ChatGPT very extensively for Q&A for 2+ years and I have hundreds of long, very technical conversations which I constantly search and refer to.
The problem (for me, anyway) is that even several megabytes worth of quality "memory" data on my profile would not allow me to migrate if it can't also confidently clone all of my chat history with it.
To be clear, this is a big enough problem that I would immediately pay low three digits dollars to have this solved on my behalf. I don't really want any of the providers to have a walled garden of all my design planning conversations, all of my PCB design conversations. Many are hundreds of prompts long. A clean break is not even remotely palatable short of OAI going full evil.
Look, I'd find it convenient for Claude to have a powerful sense of what I've been working on from conversation #1 onwards. But I absolutely refuse to bifurcate my chat history across multiple services. There is a tier list of hells, and being stuck on ChatGPT is a substantially less painful tier than needing to constantly search two different sites for what's been discussed.
I’m pretty divided on “memory”. There are times it can feel almost magical but more often than not I feel like I am fighting with the steering wheel.
Whenever I’m in a conversation and it references something unrelated (or even related) I get the “ick”. I know how context poisoning (intentional or not) works and I work hard to only expose things to the model that I want it to consider.
There have been many times that I’ve started a fresh chat as to not being along the baggage (or wrong turns) of a previous chat but then it will say “And this should work great for <thing I never mentioned in THIS chat>” and at that moment my spidey-sense tingles and I start wondering “Crap, did it come to the conclusion it did based mostly/only on the new context or did it “take a shortcut” and use context from another chat?
Like I said, I go out of my way to not “lead the witness” and so when the “witness” can peek at other conversations, all my caution is for naught.
I encourage everyone to go read the saved memories in their LLM of choice, I’ve cleaned out complete crap from there multiple times. Actually wrong information, confusing information, or one-off things I don’t want influencing future discussions.
The custom (or rather addition to the) system prompt is all I feel comfortable with. Where I give it some basic info about the coding language I prefer and the OSes that I’m often working with so that I don’t have to constantly say “actually this is FreeBSD” or “please give that to me in JS/TS instead of Python”.
The only thing that has, so far, kept me from turning off memory is that I’m always slightly cautious of going off the beaten path for something so new and moving so fast. I often want to have as close to the “stock” config since I know how testing/QA works at most places (the further off the beaten path you, the more likely you’ll run into bugs). Also so that I can experience when everyone else is experiencing (within reason).
Lastly, because, especially with LLMs, I feel like the people that over customize end up with a fragile systems. I think that a decent portion of the “N+1 model is dumber” or “X model has really gone downhill” is partially due to complicated configs (system prompts, MCP, etc) that might have helped at some point (dumber model, less capability) but are a hindrance to newer models. That or they never worked and someone just kept piling on more and more thinking it would help.
52 comments
[ 3.0 ms ] story [ 58.2 ms ] threadI find there is a virtuous cycle here where the more I use it, the more helpful it is. I fired my bookkeeper and have been using Claude with a QBO API key instead, and because it already had that context (along with other related business context), when I gave it the tax docs I gave to my CPA for 2024's taxes plus my return, and asked it to find mistakes, it determined that he did not depreciate goodwill from an acquisition. CPA confirmed this was his error and is amending my return.
Then I thought it'd be fun to see how it would do on constructing my 2024 return just from the same source docs my CPA had. First time I did it, it worked for an hour then said it had generated the return, checked it against the 2024 numbers and found they're the same. I had removed the 2024 before having it do this to avoid poisoning the context with the answers, but it turns out it had a worksheet .md file that it was using on prior questions that I had not erased (and then it admitted that it had started from the correct numbers).
In order to make sure I wouldn't have that issue again, I tried the 2024 return again, completely devoid of any historical context in a folder totally outside of my usual Claude Code folder tree. It actually got my return almost entirely correct, but it missed the very same deduction that it had caught my CPA missing earlier.
So for me, the buildup of context over time is fantastic and really leads to better results.
VSCode extension, "Please log in"
I authorize it, it creates an API key, callback. "Hello Claude, this is a test." "Please log in."
So yeah... priorities?
I switched not because I thought Claude was better at doing the things I want. I switched because I have come to believe OpenAI are a bad actor and I do not want to support them in any way. I’m pretty sure they would allow AGI to be used for truly evil purposes, and the events of this week have only convinced me further.
It's very interesting to learn more about because it challenges 1 core aspect of the economical competition : the moat.
If one can literally swap one AI service for another, then where does the valuation (and the power that comes with it) come from?
PS: I'm not interested in the service itself as I believe the side effects of large scale for-profit are too serious (and I don't mean doomdays AI takeover, I simply mean abuse of power, working conditions, downskilling, political influence as current contracts with US defense are being made, ads, ecological, etc) to be ignored.
Of course sometimes this is useful if you only use your chatbot to ask personal things like: "What should I eat today?".
But if you use it for anything else you're much better off having full control over the prompt. I can always say: "Hey btw I am german and heavily anti surveillance, what should I know about the recent anthropic DoW situation?" but with memory I lose the option of leaving out that first part.
It also showed me the difference between expectation and reality...even though these are billion dollar companies, they still haven't figured out how to make lag-free TUIs, non-Electron apps, or even respect XDG_CONFIG. The focus is definitely more on speed and stuffing these tools full of new discoveries and features right now
There's a bit of psychology around models vs. harnesses as well. You can't shake off the feeling that maybe Claude would perform better in its native harness compared to VSCode/OpenCode. Especially because they've got so many hidden skills (like the recently introduced /batch), that seem baked into the binary?
The last thing I can't figure out is computer use. Apparently all the vendors say that their models can use a mouse and keyboard, but outside of the agent-browser skill (which presumably uses playwright), I can't figure out what the special sauce is that the Cloud versions of these Agents are using to exercise programs in a VM. That is another reason why there is a switching cost between vendors.
I would assume both Claude memory and CLAUDE.md work best when they're carefully curated, only containing what you've found yourself having to repeat.
[0]: https://arxiv.org/abs/2602.11988
This way you can have Claude distill the memory as you wish.
The problem (for me, anyway) is that even several megabytes worth of quality "memory" data on my profile would not allow me to migrate if it can't also confidently clone all of my chat history with it.
To be clear, this is a big enough problem that I would immediately pay low three digits dollars to have this solved on my behalf. I don't really want any of the providers to have a walled garden of all my design planning conversations, all of my PCB design conversations. Many are hundreds of prompts long. A clean break is not even remotely palatable short of OAI going full evil.
Look, I'd find it convenient for Claude to have a powerful sense of what I've been working on from conversation #1 onwards. But I absolutely refuse to bifurcate my chat history across multiple services. There is a tier list of hells, and being stuck on ChatGPT is a substantially less painful tier than needing to constantly search two different sites for what's been discussed.
I am itching at testing claude for assembly coding and c++ to plain and simple C ports.
Must be some of the lowest switching costs I've seen which doesn't bode well for OpenAI's consumer revenues...
Whenever I’m in a conversation and it references something unrelated (or even related) I get the “ick”. I know how context poisoning (intentional or not) works and I work hard to only expose things to the model that I want it to consider.
There have been many times that I’ve started a fresh chat as to not being along the baggage (or wrong turns) of a previous chat but then it will say “And this should work great for <thing I never mentioned in THIS chat>” and at that moment my spidey-sense tingles and I start wondering “Crap, did it come to the conclusion it did based mostly/only on the new context or did it “take a shortcut” and use context from another chat?
Like I said, I go out of my way to not “lead the witness” and so when the “witness” can peek at other conversations, all my caution is for naught.
I encourage everyone to go read the saved memories in their LLM of choice, I’ve cleaned out complete crap from there multiple times. Actually wrong information, confusing information, or one-off things I don’t want influencing future discussions.
The custom (or rather addition to the) system prompt is all I feel comfortable with. Where I give it some basic info about the coding language I prefer and the OSes that I’m often working with so that I don’t have to constantly say “actually this is FreeBSD” or “please give that to me in JS/TS instead of Python”.
The only thing that has, so far, kept me from turning off memory is that I’m always slightly cautious of going off the beaten path for something so new and moving so fast. I often want to have as close to the “stock” config since I know how testing/QA works at most places (the further off the beaten path you, the more likely you’ll run into bugs). Also so that I can experience when everyone else is experiencing (within reason).
Lastly, because, especially with LLMs, I feel like the people that over customize end up with a fragile systems. I think that a decent portion of the “N+1 model is dumber” or “X model has really gone downhill” is partially due to complicated configs (system prompts, MCP, etc) that might have helped at some point (dumber model, less capability) but are a hindrance to newer models. That or they never worked and someone just kept piling on more and more thinking it would help.