I think this is going on my list of things I want to try. I have some feedback, but need to qualify it with a warning that I've barely used any AI beyond simple chat bots. This is going to be the opposite of the feedback that silentsvn gave you, meaning I have no idea what I'm talking about :-)
TLDR; You need a "how to use it" section that explains how to get information in and out of the context. That's assuming I'm not completely misunderstanding the purpose.
I started using Claude Code about a week ago, but my goal is to get something running locally that can help me get things accomplished. I'm skeptical of the claims that AI can do the work for us, but I'm interested in the idea that we can offload a bunch of cognitive load onto it freeing up brain space for the actual problems we're trying to solve. Some kind of memory system is the starting point IMO.
So here's my feedback. I skimmed the repo. You explain what it does and how it does it, but I have no idea what it does or how it does it. I think your explanations are too technical for people to understand why they'd want something like this and the example makes it look like a simple search engine. I think you need more of an explain-it-like-I'm-five approach. I might know enough to be the 5 year old in the conversation, so I'll explain a few issues I've been having and maybe you can tell me if / how your tool helps.
Most of this is in the context of using Claude Code.
I noticed the amnesia problem immediately, but expected it. I figured I'd need to take a couple of days to configure the system to remember things and adhere to my preferences, but now I realize that was wildly optimistic. Regardless, I started making a very naive system that uses markdown files with the goal of getting a better understanding of managing memory and context together. It tries to limit the current context, but it's naive. It walks a hierarchy and dumps things into the context. It's just for me to learn. I'll be happy if it helps me understand enough to pick a good tool that already exists.
The first big problem I hit was that I want what you describe as compounds, mainly chat exports, especially as I'm starting out and just want to "dump" information somewhere. I want all my chat history as I'm learning something. I had a big ah-ha moment when I asked Claude to write our conversation to a markdown file and it told me it couldn't, but offered to output a summary. I'm losing information in real time as I chat. I don't know if it's valuable or not because I don't know enough to know what I don't know.
I've been getting the most value from chatting with the AI to learn and plan things. That involves a lot of ideas, right or wrong, and I want to be able to save and retrieve those chats verbatim so I can get back to the exact same context in the future. I don't know if that's a good or bad idea, but I figure that, if I can retrieve the original context, I can always have the AI summarize it or have it help me create something more well structured once I understand the topic a bit better. I also think there's probably some value in having a future model re-evaluate that old context. For example, in the future I can start it with the current refined context (how I implemented things) and have it walk through all that old context to see if there are any novel ideas that might help to solve existing issues.
I'm assuming your spec documents are followed by the AI when working on the project. Is that right? If so, I wonder if you're underselling that by not giving an ELI5 example of how that works. For me, that's a hard problem to solve. I want a semantic search for rules the model needs to apply but I don't really want it to be semantic because they're rules that must be applied. I need to be able to ask "why isn't the tool following my docker compose spec" and need a deterministic way to answer that. I think your project does that.
Maybe I'm simply lacking knowledge and should be able to understand why I need this kind of tool and, more importantly, how it maps to context management (assuming that's what it does).
I'll give you an analogy, at least that applies to me. Your "how it works" section is like going to driver training and having the instructor start explaining how the car's engine and transmission are built. People like me need it dumbed down; "Push the gas and turn the wheel. It's faster than your bicycle."
Maybe I'm not the target audience yet, but maybe I am. I'm already convinced that AI with good memory management is useful. I'm also unwilling to build that memory using a commercial system like Claude or ChatGPT. It's vendor lock-in on the level of getting a lobotomy if you lose access to that system and I don't think people are doing a good job of assessing that risk.
I'm going to finish building my own crappy memory system and then yours is going to be the first real system I try. Thanks for sharing it.
I wonder how it works and how much heavy lifting "supervising" is doing. Whenever I try to use AI, the outcome is about the same.
It's good at non-critical things like logging or brute force debugging where I can roll back after I figure out what's going on. If it's something I know well, I can coax a reasonable solution out of it. If it's something I don't know, it's easy to get it hallucinating.
It really goes off the rails once the context gets some incorrect information and, for things that I don't understand thoroughly, I always find myself poisoning the context by asking questions about how things work. Tools like the /ask mode in Aider help and I suspect it's a matter of learning how to use the tooling, so I keep trying.
I'd like to know if AI is writing code their best developers couldn't write on their own or if it's only writing code they could write on their own because that has a huge impact on efficiency gains, right? If it can accelerate my work, that's great, but there's still a limit to the throughput which isn't what the AI companies are selling.
I do believe there are gains in efficiency, especially if we can have huge contexts the AI can recall and explain to us, but I'm extremely skeptical of who's going to own that context and how badly they're going to exploit it. There are significant risks.
If someone can do the work of 10 people with access to the lifetime context of everyone that's worked on a project / system, what happens if that context / AI memory gets taken away? In my opinion, there needs to be a significant conversation about context ownership before blindly adopting all these AI systems.
In the context of Spotify in this article, who owns the productivity increase? Is it Spotify, Anthropic, or the developers? Who has the most leverage to capture the gains from increasing productivity?
There's no definitive way to tell if some code is written by AI or not; thus their statement doesn't have to be true. Usage of AI itself is nebulous, it could mean anything from OpenClaw-style automated agents to someone prompting an LLM via chat to an engineer manually writing code after wasting hours trying to get an agent to do it (that still counts as "usage", even if not ultimately productive).
The market currently rewards claims of usage of AI, so everyone is claiming to be using AI. There is no way to prove one way or another, and the narrative will be changed down the line once the market swings.
When it comes to productivity claims, I have yet to see it. If AI is truly providing significant, sustained productivity improvements across the software development lifecycle I'd expect products to be better, cheaper, or get developed faster (all of which happened with other industrial breakthroughs). I do not see that in software at large and especially not in Spotify's case.
Existing large market products and services most likely will not bring down the price of their goods and services even if AI reduces the cost to produce.
There is a saying in sales, "Money left on the table". People are already willing to pay current market value prices so they will keep doing so. Why decrease your profits when people are already willing to pay? Keeping the same pricing would maximize profits and stock, a bonus.
New companies that start producing alternative solutions would most likely be the ones to utilize the reduction in cost for producing with AI. It would be their way to get a foot in the door and start building market.
Organizations that focus on providing a good or services to the vulnerable would be an exception. Some companies that market on cost would be in that mix too.
Fair, but in this case you'd still expect LLMs to be a massive boon to startups wanting to compete with the established players?
I'm not seeing any of that happening either, which I'd expect if the AI productivity claims were true and anyone could just vibe-code an entire sustainable business.
Imagine having 20 years of context / memories and relying on them. Wouldn't you want to own that? I can't imagine pay-per-query for my real memories and I think that allowing that for AI assisted memory is a mistake. A person's lifetime context will be irreplaceable if high quality interfaces / tools let us find and load context from any conversation / session we've ever had with an LLM.
On the flip side of that, something like a software project should own the context of every conversation / session used during development, right? Ideally, both parties get a copy of the context. I get a copy for my personal "lifetime context" and the project or business gets a copy for the project. However, I can't imagine businesses agreeing to that.
If LLMs become a useful tool for assisting memory recall there's going to be fighting over who owns the context / memories and I worry that normal people will lose out to businesses. Imagine changing jobs and they wipe a bunch of your memory before you leave.
We may even see LLM context ownership rules in employment agreements. It'll be the future version of a non-compete.
Whoever is paying for it? If you've got personal stuff you'd keep it in your own account (or maintain it independently), separate from your work account.
I'd be interested to hear your thoughts on having a PWA vs regular mobile apps since it looks like you started with a PWA, but are moving to regular apps. Is that just a demand / eyeballs thing or were there technical reasons?
I've used https://historio.us since 2011 and still pay for it to keep access to all the pages I've archived over the years. The price has been kept low enough that I can't bring myself to cancel it even though I've been using self-hosted https://archivebox.io/ for the last few years.
I always include an archived link whenever I reference something in documentation. That's my main use at the moment.
However, I also feel like I've gotten a lot of really good value when trying to learn a new development topic. Whenever I find something that looks like it might be useful, I archive it and, because everything is searchable, I end up with a searchable index of really high quality content once I actually know what I'm doing.
I find it hard to rediscover content via web search these days and there's so much churn that having a personal archive of useful content is going to increase in value, at least in my opinion.
How much space is the self-hosted solution taking? I've been meaning to try and find a better way to look through my bookmarks since no browser is capable of doing that properly it seems.
> .com itself is under jurisdiction of USA and operated by Verisign
Barely. The NTIA gave up all their leverage over .com in 2018. The only thing the US can do at this point is let the cooperative agreement auto-renew to limit price increases.
I wouldn't be surprised if the US withdrew from the agreement altogether at this point. Then .com would fall under the joint control of ICANN and Verisign.
It's going to be interesting to see what they do. One of the core arguments when claiming the domain industry enjoys a competitive market is that switching costs are bearable and that switching TLDs is an option if registries increase prices too much.
So ICANN has a non-trivial choice to make. Either they maintain the position that switching costs are bearable and let .io disappear, or they admit that TLD switching is impossible and save .io, which will make it hard to argue the threat of (registrants) TLD switching keeps the industry competitive.
> They can't target popular domains for discriminatory pricing.
That's not completely accurate. Section 2.10c of the base registry agreement says the following in relation to the uniform pricing obligations:
> The foregoing requirements of this Section 2.10(c) shall not apply for (i) purposes of determining Renewal Pricing if the registrar has provided Registry Operator with documentation that demonstrates that the applicable registrant expressly agreed in its registration agreement with registrar to higher Renewal Pricing at the time of the initial registration
Most registrars have blanket statements in their registration agreement that say premium domains may be subject to higher renewal pricing. For registry premium domains, there are no contractual limits on pricing or price discrimination. AFAIK, the registries can price premium domains however they want.
You omitted key portions of that section. Here's the full quote (emphasis added):
> The foregoing requirements of this Section 2.10(c) shall not apply for (i) purposes of determining Renewal Pricing if the registrar has provided Registry Operator with documentation that demonstrates that the applicable registrant expressly agreed in its registration agreement with registrar to higher Renewal Pricing at the time of the initial registration of the domain name following clear and conspicuous disclosure of such Renewal Pricing to such registrant
Furthermore:
> The parties acknowledge that the purpose of this Section 2.10(c) is to prohibit abusive and/or discriminatory Renewal Pricing practices imposed by Registry Operator without the written consent of the applicable registrant at the time of the initial registration of the domain and this Section 2.10(c) will be interpreted broadly to prohibit such practices
Yes, premium domains can be priced higher, but the Renewal Pricing has to be "clear and conspicuous" to the registrant at the time of initial registration. Are you aware of any litigation related to this?
The exact pricing isn’t disclosed. All they do is tell you the price will be “higher”. Anyone registering a premium domain is getting higher than uniform renewal pricing, so whatever they’re doing right now is considered adequate and that’s just generic ToS in the registration agreement AFAIK.
It sounds like you think I’m being deceptive. Do you know about any registry premium domains where someone has a contractually guaranteed price?
Also, based on my own anecdotal experience, ICANN doesn’t interpret 2.10c broadly and they allow the registries to push the boundaries as much as they want.
I think that once you have domains as an identity, you can solve a lot of problems with the idea of 'just add money'. If $1000 gets me a gold check mark, it changes the economics of impersonation. Is it worth it to spend $1000 to get a gold check mark on 'goog1e.com' if a brand monitoring system is going to get that moderated out of existence in a couple of hours?
That's also why domain verification systems need to have continuous re-validation with more frequent re-validation for new identities. For example, if '@goog1e.com' is a new identity, it should be re-validated after 1h, 4h, 8h, 16h (up to a maximum). Additionally, you could let other validated users with aged accounts trigger a re-validation (with shared rate limits for a target domain).
The great thing about domains is that those of us that are good faith participants can build a ton of value on them and that value can be used as a signal for trustworthiness. The hard part is conveying that value to regular users in a way that's simple to understand.
We could also have systems that use some type of collateral attestation. For example, if I donate $1000 to the EFF, maybe I could attribute that donation to my domain 'example.com' and the EFF could attest to the fact that I've spent $1000 in the name of 'example.com'.
You probably have to gate that though some type of authority, but I can imagine a system where domain registrars could do that. I would love to buy reputation from my registrar by donating money to charity.
In the latter case, if you are the EFF, or any other recognized charity (and if you allow a lot of charities that's a lot of people) you can assign a trillion dollars to any domain you like, which is usually cited as a reason to avoid this type of system.
And if the EFF turns bad in the future you can't get a verification badge without supporting bad guys.
This is always true any time you have more than 1 human involved. People can always become corrupt and dishonest, and no technological solution will solve that.
The platform owners have spent two decades de-emphasizing domains, so it's not too surprising that most people struggle to understand how they work. I think that can change with education and awareness if domains as identity start to catch on. It just takes time.
For now, I think wider adoption of things like DomainConnect [1] would make a difference. It works really well to set up an MS365 account with DNS hosted at Cloudflare, but it would need a workflow that supports sending requests to your DNS admin rather than assuming everyone is a DNS admin.
> A lot of people do not want to look at and understand domain names, instead they want to see a name and a check mark. They want a central authority to tell them who is trustworthy and who is not.
I think 'trustworthy' is a key word there and would add that I think a lot of regular people conflate identity verification with moderation. It's important to keep those separate because as soon as an identity system becomes a moderation system, it's worthless.
That's what makes domains so great for identity, especially with the way the AT protocol works. It helps to create a clear separation between identity verification and moderation. Moderation is much harder than identity verification, so having a clear line between the two should make it easier to develop technical systems that perform identity verification.
For pure identity verification, I think BIMI [2] is sitting on a solution they don't even realize they have. They're too tunnel visioned on email verification, but the system they've built with VMC (verified mark certificates) works as a decentralized system of logo verification. For example, I can tell you this logo [3] is trademarked and owned by 'cnn.com' and I can do it via technical means starting with the domain name:
dig default._bimi.cnn.com TXT
Seeing a 3rd party URL in the TXT value makes me think the implementation is weak since that would be better as a CNAME pointing to a TXT record managed by a 3rd party, but I've never looked into the details enough to know if it'll follow CNAMEs (like ACME or DKIM do).
Also, the VMCs are only good for high value brands because CNN is paying DigiCert $1600 / year for the certificate, but, since it's just PKI, it allows anyone to put up that logo with a verified badge on the @cnn.com identity. A more accurate badge would be the registered trademark symbol [4].
Even though that only works for high value brands that own a logomark, it works extremely well and would be a great start to a system that's easier for the average person to understand because logos are a simpler concept than something abstract like domains and no one is spending the time and effort needed to get a fake VMC (if it's even possible).
The Bluesky implementation for domain verification has a long way to go though. It's very naive at the moment and doesn't even do a proper job of dealing with changes in domain ownership. In fact, almost everyone doing domain validation is doing it wrong because very few implementation do re-validation from what I've seen.
TLDR; You need a "how to use it" section that explains how to get information in and out of the context. That's assuming I'm not completely misunderstanding the purpose.
I started using Claude Code about a week ago, but my goal is to get something running locally that can help me get things accomplished. I'm skeptical of the claims that AI can do the work for us, but I'm interested in the idea that we can offload a bunch of cognitive load onto it freeing up brain space for the actual problems we're trying to solve. Some kind of memory system is the starting point IMO.
So here's my feedback. I skimmed the repo. You explain what it does and how it does it, but I have no idea what it does or how it does it. I think your explanations are too technical for people to understand why they'd want something like this and the example makes it look like a simple search engine. I think you need more of an explain-it-like-I'm-five approach. I might know enough to be the 5 year old in the conversation, so I'll explain a few issues I've been having and maybe you can tell me if / how your tool helps.
Most of this is in the context of using Claude Code.
I noticed the amnesia problem immediately, but expected it. I figured I'd need to take a couple of days to configure the system to remember things and adhere to my preferences, but now I realize that was wildly optimistic. Regardless, I started making a very naive system that uses markdown files with the goal of getting a better understanding of managing memory and context together. It tries to limit the current context, but it's naive. It walks a hierarchy and dumps things into the context. It's just for me to learn. I'll be happy if it helps me understand enough to pick a good tool that already exists.
The first big problem I hit was that I want what you describe as compounds, mainly chat exports, especially as I'm starting out and just want to "dump" information somewhere. I want all my chat history as I'm learning something. I had a big ah-ha moment when I asked Claude to write our conversation to a markdown file and it told me it couldn't, but offered to output a summary. I'm losing information in real time as I chat. I don't know if it's valuable or not because I don't know enough to know what I don't know.
I've been getting the most value from chatting with the AI to learn and plan things. That involves a lot of ideas, right or wrong, and I want to be able to save and retrieve those chats verbatim so I can get back to the exact same context in the future. I don't know if that's a good or bad idea, but I figure that, if I can retrieve the original context, I can always have the AI summarize it or have it help me create something more well structured once I understand the topic a bit better. I also think there's probably some value in having a future model re-evaluate that old context. For example, in the future I can start it with the current refined context (how I implemented things) and have it walk through all that old context to see if there are any novel ideas that might help to solve existing issues.
I'm assuming your spec documents are followed by the AI when working on the project. Is that right? If so, I wonder if you're underselling that by not giving an ELI5 example of how that works. For me, that's a hard problem to solve. I want a semantic search for rules the model needs to apply but I don't really want it to be semantic because they're rules that must be applied. I need to be able to ask "why isn't the tool following my docker compose spec" and need a deterministic way to answer that. I think your project does that.
Maybe I'm simply lacking knowledge and should be able to understand why I need this kind of tool and, more importantly, how it maps to context management (assuming that's what it does).
I'll give you an analogy, at least that applies to me. Your "how it works" section is like going to driver training and having the instructor start explaining how the car's engine and transmission are built. People like me need it dumbed down; "Push the gas and turn the wheel. It's faster than your bicycle."
Maybe I'm not the target audience yet, but maybe I am. I'm already convinced that AI with good memory management is useful. I'm also unwilling to build that memory using a commercial system like Claude or ChatGPT. It's vendor lock-in on the level of getting a lobotomy if you lose access to that system and I don't think people are doing a good job of assessing that risk.
I'm going to finish building my own crappy memory system and then yours is going to be the first real system I try. Thanks for sharing it.
reply