As long as there's no solution to the long-term memory problem, we will have a "country of geniuses in a data center" that are all suffering from anterograde amnesia (movie: Memento), which requires human hand-holding.
I have experimented with a lot of hacks, like hierarchies of indexed md files, semantic DBs, embeddings, dynamic context retrieval, but none of this is really a comprehensive solution to get something that feels as intelligent as what these systems are able to do within their context windows.
I am als a touch skeptical that adjusting weights to learn context will do the trick without a transformer-like innovation in reinforcement learning.
I've used open claw (just for learning, I agree with the author it's not reliable enough to do anything useful) but also have a similar daily summary routine which is a basic gemini api call to a personal mcp server that has access to my email, calendar etc. The latter is so much more reliable. Open claw flows sometimes nail it, and then the next day fails miserably. It seems like we need a way to 'bank' the correct behaviours - like 'do it like you did it on Monday'. I feel that for any high percentage reliability, we will end up moving towards using LLMs as glue with as much of the actual work as possible being handed off to MCP or persisted routine code. The best use case for LLMs currently is writing code, because once it's written, tested and committed, it's useful for the long term. If we had to generate the same code on the fly for every run, there's no way it would ever work reliably. If we extrapolate that idea, I think it helps to see what we can and can't expect from AI.
This is interesting. I haven't used OpenClaw but I set up my own autonomous agent using Codex + ChatGPT Plus + systemd + normal UNIX email and user account infrastructure. And it's been working great! I'm very happy with it. It's been doing all kinds of tasks for me, effectively as an employee of my company.
I haven't seen any issues with memory so far. Using one long rolling context window, a diary and a markdown wiki folder seems sufficient to have it do stuff well. It's early days still and I might still encounter issues as I demand more, but I might just create a second or third bot and treat them as 'specialists' as I would with employees.
I did (using Claude Code) something that sounds very similar to this. It’s a bunch of bootstrapped Unix tools, systemd units, and some markdown files. Two comments:
- I suspect that in this moment, cobbling together your own simple version of a “claw-alike” is far more likely to be productive than a “real” claw. These are still pretty complex systems! And if you don’t have good mental models of what they’re doing under the hood and why, they’re very likely to fail in surprising, infuriating, or downright dangerous ways.
For example, I have implemented my own “sleep” context compaction process and while I’m certain there are objectively better implementations of it than mine… My one is legible to me and therefore I can predict with some accuracy how my productivity tamagotchi will behave day-to-day in a way that I could not if I wasn’t involved in creating it.
(Nb I expect this is a temporary state of affairs while the quality gap between homemade and “professional” just isn’t that big)
- I do use mine as a personal assistant, and
I think there is a lot of potential value in this category for people like me with ADD-style brains. For whatever reason, explaining in some detail how a task should be done is often much easier for me than just doing the task (even if, objectively, there’s equal or higher effort required for the former). It therefore doesn’t do anything I _couldn’t_ do myself. But it does do stuff I _wouldn’t_ do on my own.
Right - I think email is a much better UI than Slack or WhatsApp or Discord for that reason. It forces you to write properly and explain what you want, instead of firing off a quick chat. Writing things down helps you think. And because coding harnesses like Codex are very good at interacting with their UNIX environments but are also kinda slow, email's higher latency expectations are a better fit for the underlying technology.
Two categories: actual useful work for the company, and improving the bot's own infrastructure.
Useful work includes: bug triage, matching up external user bug reports on GitHub to the internal YouTrack, fixing easy looking bugs, working on a redesign of the website. I also want to extend it to handling the quarterly accounting, which is already largely automated with AI but I still need to run the scripts myself, preparing answers to support queries, and more work on bug fixing+features. It has access to the bug tracker, internal git and CI system as if it were an employee and uses all of those quite successfully.
Meta-work has so far included: making a console so I can watch what it's doing when it wakes up, regularly organizing its own notes and home directory, improving the wakeup rhythm, and packaging up its infrastructure to a repeatable install script so I can create more of them. I work with a charity in the UK whose owner has expressed interest in an OpenClaw but I warned him off because of all the horror stories. If this experiment continues to work out I might create some more agents for people like him.
I'm not sure it's super useful for individuals. I haven't felt any great need to treat it as a personal assistant yet. ChatGPT web UI works fine for most day to day stuff in my personal life. It's very much acting like an extra employee would at a software company, not a personal secretary or anything like that.
It sounds like our experience differs because you wanted something more controlled with access to your own personal information like email, etc, whereas I gave "Axiom" (it chose its own name) its own accounts and keep it strictly separated from mine. Also, so far I haven't given it many regular repeating tasks beyond a nightly wakeup to maintain its own home directory. I can imagine that for e.g. the accounting work we'd need to do some meta-work first on a calendar integration so it doesn't forget.
I’m doing this exact same thing in my solo saas company, except with Cursor’s Cloud Agents. I can kick them off from web, slack, linear, or on a scheduled basis, so I’m doing a lot of the same things as you. It’s just prompts on a cron, with access to some tools and skills, but super useful.
That unreliability was why I gave up on OpenClaw. I tried hard to give it very simple tasks but it had a high degree of failure. Heartbeats and RAG are lightyears away from where they need to be. I'm not sure if this can be overcome using an application layer right now, but I trust that many people are trying, and I'm eager to see what emerges in the next year. In the mean time I know that they're working very hard on continuous learning - real-time updates to weights and parametric knowledge. It could be that in a year or so, we can all have customised models.
That would be great if that comes to fruition. Investing in a model with weights updates would be like investing in employee training, rather than just giving the same unreliable employee more and more specific instructions.
You're right to be skeptical. Without a way to actually implement how the human brain processes experiences into a consolidated memory, we won't be able to solve the long term memory problem at all. Not with the current technology.
An LLM context is a pretty well extended short term memory, and the trained network is a very nice comprehensive long term memory, but due to the way we currently train these networks, an LLM is just fundamentally not able to "move" these experiences to long term, like a human brain does (through sleep, among others).
Once we can teach a machine to experience something once, and remember it (preferably on a local model, because you wouldn't want a global memory to remember your information), we just cannot solve this problem.
I think this is probably the most interesting field of research right now. Actually understanding in depth how the brain learns, and figuring out a way to build a model that implements this. Because right now, with backtracking and weight adjustments, I just can't see us getting there.
I think if we want to build on what we have, instead of compaction at the end of the context window, the LLM would have to 'sleep', i.e. adjust its weights, then wake up with the last bits of the old context window in the new one, and have a 'feel' for what it did before through the change in weights. I just sense it's not that simple to get there, because simply updating the weights based on a single context sample risks degrading the weights of the whole network.
I like the idea of using small local model (or several) for tackling this problem, like low rank adaptation, but with current tech, I still have to piece this together or the small local models will forget old memories.
Sleep would probably be a part of the equation for consolidating , but there's still the question of how exactly does the brain process the information during sleep in a way that it permanently consolidates the information.
It's not how an llm can work right now, it needs too much iterations & a much bigger dataset than what we can work with. A single time experiencing something and we can remember it. That's orders of magnitude more efficient than an LLM right now can achieve.
Couldn't fitting solve the problem? That's what companies do: take a model as a base and train it on the specific data long enough so that it prefers the new data.
Overfitting may be a thing but for personal use, I may want to have it work as I expected, every time.
> I think this is probably the most interesting field of research right now. Actually understanding in depth how the brain learns, and figuring out a way to build a model that implements this.
This field of research has been around for decades, so who's to say when there'll be a breakthrough.
In fact, LLMs are great despite our very limited understanding, and not because we had some breakthrough about the human brain.
Exactly. It's been around so long and we still don't know how to mimic it.
The way an llm learns is a very interesting way of doing it, but it sure isn't what the brain is doing.
But it's indisputable.. We can get enormous results with this technique. It's just probably not the way forward for faster learning to remediate the issue of context loss.
Why does a language model have to be monolithic? I think retraining a model is expensive (relatively speaking). Is there some way to bolt on specialization?
It's kind of fascinating that everyone is trying to build a Chinese Room agent with stateless models, since we don't know how to produce a stateful model with continuous, incremental training.
It's like spontaneous implemention of thought experiments from yesteryear. I wonder if all this product-focused experimentation will accidentally impact philosophy of mind after all...
I agree. A key to human intelligence is our ability to adjust our weights in real-time. All knowledge becomes parametric knowledge - the knowledge stored inside the model. RAG is a messy workaround which requires making assumptions about what is needed to load from external sources before it is clear what is needed. Agentic loops can go some way to overcome this, but they are resource intensive, slow, prone to mistakes and deviations, and far less accurate. The secret sauce of an LLM is the vectorised weights. RAG is like putting a 1990s Honda Civic engine into a Ferrari. You can do it, but the result is quite terrible.
I think we will eventually end up with models which can be individually trained and customised on regular schedules. After that, real-time.
From my perspective there are some people that have never built real processes in their life that enjoy having some processes now. But agent processes are less reliable slower and less maintenable then a process that is well-defined and architectured and uses llm’s only where no other solution is sufficient. Classification, drafting, summarizing.
I’ve had a Whatsapp assistant since 2023, jailbraked as easy assistant. Only thing I kept using is transcription.
https://github.com/askrella/whatsapp-chatgpt was released 3 years ago and many have extended it for more capabilities and arguably its more performant than Openclaw as it can run in all your chat windows. But there’s still no use case.
I like to experiment with AI flows to make iteration quicker, then once something work investing in is found, back up and build something that's actually repeatable.
Same thing could be said with SKILL.md yet they are highly useful...
Yes you can automate via scripting, but interacting with a process using natural language because every instance could be different and not solid enough to write a spec for, is really handy.
tl;dr: there's a place for "be liberal in what you receive and conservative in what you send", but only now have LLMs provided us with a viable way to make room for "be loosey goosey with your transput"
I understand but there still is usually 80-95% of the skill flow that you can script out that is repeated. Script it out and
simplify your skill, make it more stable, and provide more opportunity to scale it up or down i.e use stronger or weaker models if need be. We should be scripting and forming process first then seeing where we can put AI after that.
I find that it's usually management that ask for such things "because AI".
I mean using AI is a great way to interpret a query, determine if a helper script already exists to satisfy it, if not invoke a subagent to write a new script.
Problem with your "script" approach is how does that satisfy unknown/general queries? What if for one run you want to modify the behavior of the script?
Fully agree with your comment regarding real processes. Being a Six Sigma Black Belt, studying processes and reducing the errors is critical.
The Openclaw processes at the moment scare me.
I'm using openclaw as a personal development bot, which is pretty useful. It pings me throughout the day using crons to complete tasks and follows up on them. But aside from that, it is a very unreliable piece of software. I'm constantly having to fix it, or track down correct configurations. It can just decide to randomly edit it's own config, uses incorrect json keys and then the whole thing is dead. Or it blows through it's context and doesn't know to compact. Then it's just stuck. I can't wait till it matures or something more reliable comes along.
It would’ve happened eventually anyway, but OpenClaw is basically what kickstarted the beginning of the end of token subsidies. It’s a almost begging to be used wastefully. And agents would miss and lose nothing without it. It’s devoid of a reason to exist.
I actually quite enjoy the OpenClaw. Although the recent CC crackdown has caused me to try different LLM providers which aren't that reliable but anyways, here are few things I do with it, in all separate groups.
* Telegram Health Group, created an agent to help me track sleep, recommend my supplements based on my location, remind me in the morning and evening to monitor my food. I send it images of what I eat and it keeps track of it.
* Telegram Career Group, I randomly ask it to find certain kind of job posts based on my criteria. Not scheduled, only when I like to.
* Telegram Coder Group, gave it access to my github account. It pulls, runs tests and merges dependabot PRs in the mornings. Tells me if there are any issues. I also ask it to look into certain bugs and open PRs while I'm on the road.
* Telegran News Group, I gave it a list of youtube videos and asked it to send me news every day at 10am similar to the videos.
So far, it's a super easy assistant taking multiple personas.
But it's getting a bit painful without CC subscription
I built a special belief-based system recently for my own agent harnesses instead of some similarity based fact storage stuff... which falls flat once conflicting data points enter the system and just increase LLM confusion and make it do weird things. this means learning over time works a bit more like humans do - superseding old beliefs and reconciliating stuff cleanly over time. Also including the building blocks to have a subagent managing it autonomously (with tools/skills/soul). works quite well and very fast given its pure nodejs+sqlite and doesn't eat tokens like crazy or needs any thirdparty embeddings solution. maybe have a look.
Why aren't databases the solution to many memory problems? Maybe this is a naive question:
For example, for the invitations in the OP: Have Openclaw write incoming rsvps to a database, probably a flat file here, and use the db as persistent memory: OpenClaw can compose outgoing update emails based on the database. Don't even suggest to OpenClaws that it try to remember the rsvps - its job is just writing to and reading from a database, and composing emails based on the latter. ?
Does that violate the experiment, by using some tool in addition to OpenClaw?
I partly identify with the article. While I don't use OpenClaw itself, I hacked together my own small Claude-in-a-loop/cronjob, and it seems we're all getting our morning briefings and personalized morning podcasts now.
The other common use case seems to be kicking off an automated Claude session from an email / voicetext / text / Telegram, and getting replies back. I'm emailing Claude throughout the day now, and sometimes it's useful to just forward an email to Claude and ask it to handle the task within it for me.
But I think many people criticizing the various Claws are missing out on the cronjob aspect. There's value in having your AI do work automatically while you're asleep. You don't even need OpenClaw for that, just a cronjob that runs claude -p in the early morning. If you give your AI enough context about yourself, you get to a point where it just independently works on things for you, and comes to you with suggestions. It doesn't need to be specifically prompted. The environment of data it can access is its own context, its own prompt. With that, it can sometimes be surprising and spooky what you wake up to, without being directly prompted.
Give it enough context, long term memory, and ability to explore all of that, and useful stuff emerges.
I've had a crack at this problem in Agent Kanban for VS Code (https://github.com/appsoftwareltd/vscode-agent-kanban). The core idea is that you converse with the agent in a markdown task file in a plan, todo, implement flow, and that I have found works really well for long running complex tasks, and I use this tool every day. But after a while, the agent just forgets to converse in the task file. The only way to get it to (mostly) reliably converse in the task file is to reference the task file and instructions in AGENTS.md. There is support for git work trees and skipping commits of the agents file so as not to pollute the file with the specific task info. There is also an option for working without work trees, but in this flow I had to add chat participant "refresh" commands to help the agent keep it's instructions fresh in context. It's a problem that I believe will slowly get better as better agents appear, and get cheaper to use, because general LLM capability is the key differentiator at the moment.
If you look at my comment history, you'll see what seems to be someone defending OpenClaw (even though I stopped using it).
I have some issues with the article, but I agree with some of the conclusions: It's great tinkering with it if you have time to spare, but not worth using weeks of your time trying to get a perfect setup. It's just not that reliable to use up so much of your time.
I will say, it's still amongst the best tools to do a variety of tasks. Yes, each one of those could be done with just a coding agent, but I found it's less effort to get OpenClaw to do it than you writing something for each use case.
Very honest question: One of the use cases I had with OpenClaw that I'm missing now that I don't use it: I could tell it (via Telegram) to add something to my TODO list at home while I'm in the office. It would call a custom API I had set up that adds items to my TODO list.
How can I replicate this without the hassle of setting up OpenClaw? How would you do it?
(My TODO list is strictly on a home PC - no syncing with phone - by design).
(BTW, the reason I stopped using OpenClaw is boring: My QEMU SW stopped working and I haven't had time to debug).
I’ve had some success with claude cli agents at some scale with a memory architecture - but it roughly reads like a massive index, where it crawls through a trail of breadcrumbs to piece together all the info it needs to do a task. It’s fairly tedious to maintain, and it’s always a battle maintaining reasonable context size and token spend.
I’d say it’s like 85% reliable on any given task, and since I supervise it, this is good enough for me. But for something to be useful autonomously, that number needs to be several 9’s to be useful at all, and we’re no world near that yet.
I’m currently watching someone trying and failing to roll openclaw out at scale in an org and they believe in it so much it’s very difficult to convince them even with glaring evidence staring them in the face that it will not work
Some of this stuff is starting to look like technologies that worked, looked promising, but were at best marginally useful, such as magnetohydrodynamic generators, tokamaks, E-beam lithography, and Ovonics.
I primarily "only" use it as a run-manager that can spin up another agent in a tmux which I can then join by ssh on my cell phone. Then, I can monitor the work from my cell phone and choose to either directly interact with the tmux pane or else just message my openclaw agent to do it for me. That right there is the only "killer" app I've found for it. I do also use it to post to my x.com account and that's also pretty useful. Neither of these uses assume any super long context over time will be retained. But, to me, the run-manager use case is pretty great.
This is essentially a 140-IQ analyst you can text "please fix" at 3:00 AM without the shit you have to deal with when hosting your stuff locally. The context awareness across your text messages, emails, and between chats is not half bad; just wish that the "computer use" feature set was fully locked in so I can automate half the apps on my phone.
Openclaw is unreliable.
I had it running for a few months. It uses up a lot of resources and doesn't provide any benefit other than being able to chat with it via other methods than tui.
I dunno - my boss has deployed a couple of claw agents that are pretty good at doing SWE and SRE work. They’re available for the whole company to use, and they save us a ton of time. Pretty decent use case! Personally I haven’t found claw agents replace anything really for personal use outside of commercial tools I’d pay for to handle scheduling and stuff, but I also haven’t tried / trusted too many new use cases outside of that cron / daily briefing or some family schedules.
It is an interesting take. I think this is mainly early adoption pains though. This stuff is moving so fast that if you say 'it isn't useful because X isn't good enough' then just wait a month and X will be good enough to find Y as the blocker (or no blockers are left and it truly does become useful). Soon we will see this hooked into the home assistant world well combined with local and remote compute and then we are likely to see real movement.
Author basically admitting to having a boring outlook on life IMO. Sure maybe not for work but there's tons of things that suck the life out of your limited time, having a tool not just OpenClaw is one way to not bend to the will of BigCo for whatever thing you want to do, need a 3D model? Need something summarized or need control of something which the manufacturer forbids? All can be done without spending entire weekends for.
The hype around OpenClaw is a bit confusing but I think I figured it out. For most coders, Claude Code in the terminal was an important event. Letting it access code and change files directly. For normal users, they didn’t see the power is that.
OpenClaw runs Pi in a terminal and exposes the chat thru Telegram or any chatting app. This gave the ah-ha moment to non-coders that coders had had for 6+ months prior.
Good to know that I'm not alone. I now use it for music recommendations (not so great) and keeping track of restaurants I want to try (really good at this but so are a lot of other apps).
It can integrate apis for you on the fly. That’s one of the biggest usecases IMO. Combine that with skills, cron, and sub-agents, and you get a lot of power there.
I'm not sure what these people who have strong opinions like this think Openclaw is, but to me, it's a product with 1) a somewhat easy to setup prompt passing wrapper that can span many channels like Telegram, Whatsapp etc. 2) A (at least optimistically) plug-n-play, configurable architecture to wake up to events (cron entries, webhooks etc.) and fire up agents in order to get 'proactive' behavior, with the flexibility to integrate models from a gazillion providers. Pretty much everything else it's bundled with is general purpose tooling that does or could easily exist in any other agentic tool.
It's a rather simple framework around an LLM, which actually was a brilliant idea for the world that didn't have it. It also came with its own wow effect, ("My agent messaged me!") so I consider some of the hype as justified.
But that's pretty much it. If you can imagine use cases that might involve emailing an LLM agent and get responses that share context with other channels and resources of yours, or having the ability to configure scheduled/event-based agent runs, you could get some use out of having an Openclaw setup somewhere.
I find the people who push insanity like "It came alive and started making money for me" and the people who label it utterly, completely useless (because it has the same shortcomings as every other LLM-based product) like Mr. "I've Seen Things. Here's the Clickbait" here, rather similar. It's actually hard to believe they know what they're talking about or that they believe what they're writing.
Sure, anything it does can be done better with specialized tooling. If you know that tooling.
The memory thing sounds like an implementation limit rather than something fundamentally unsolvable. Just experiment with different ways of organizing state until something works?
IMHO, the biggest problem with OpenClaw and other AI agents is that the use-cases are still being discovered. We have deployed several hundred of these to customers and I think this challenge comes from the fact that AI agents are largely perceived as workflow automation tools so when it comes to business process they are seen as a replacement for more established frameworks.
They can automate but they are not reliable. I think of them as work and process augmentation tools but this is not how most customers think in my experience.
However, here are a several legit use-case that we use internally which I can freely discuss.
There is an experimental single-server dev infrastructure we are working on that is slightly flaky. We deployed a lightweight agent in go (single 6MB binary) that connects to our customer-facing API (we have our own agentic platform) where the real agent is sitting and can be reconfigured. The agent monitors the server for various health issues. These could be anything from stalled VMs, unexpected errors etc. It is firecracker VMs that we use in very particular way and we don't know yet the scope of the system. When such situations are detected the agent automatically corrects the problems. It keeps of log what it did in a reusable space (resource type that we have) under a folder called learnings. We use these files to correct the core issues when we have the type to work on the code.
We have an AI agent called Studio Bot. It exists in Slack. It wakes up multiple times during the day. It analyses our current marketing efforts and if it finds something useful, it creates the graphics and posts to be sent out to several of our social media channels. A member of staff reviews these suggestions. Most of the time they need to follow up with subsequent request to change things and finally push the changes to buffer. I also use the agent to generate branded cover images for linkedin, x and reddit articles in various aspect ratios. It is a very useful tool that produces graphics with our brand colours and aesthetics but it is not perfect.
We have a customer support agent that monitors how well we handle support request in zendesk. It does not automatically engage with customers. What it does is to supervise the backlog of support tickets and chase the team when we fall behind, which happens.
We have quite a few more scattered in various places. Some of them are even public.
In my mind, the trick is to think of AI agents as augmentation tools. In other words, instead of asking how can I take myself out of the equation, the better question is how can I improve the situation. Sometimes just providing more contextually relevant information is more than enough. Sometimes, you need a simple helper that own a certain part of the business.
My biggest issue with OpenClaw is everyone talks about doing things with it but doesn't explain what it actually is doing.
First of all is not an LLM, you're beholden to an api or local llm limitations. Second of all it's always calendars, email replies, summarizing.
You do not need an LLM for that, and an LLM doesn't make it easier either. It sounds like executive cosplay, not productivity. Everything I see people talking about that's actually productive, it's doing probabilistically when deterministic tools already exist and have for in some cases over 20 years.
You don't need an LLM to put a meeting on a calendar, that's literally two taps with your phone or a single click in gmail. Most email services already have suggestions already built in. Emails have been summarized for 10 years at this point. If you're so busy you need this stuff automated, you probably have an assistant, or you're important enough that actually using general intelligence is critical to being successful at all.
The idea of getting an LLM email response sounds great for someone who has never worked a job in their life.
This comment section is full of llm writen responses too, to the point where its absurd. Noticing how most of them just talk in circles like "But I think many people criticizing the various Claws are missing out on the cronjob aspect. There's value in having your AI do work automatically while you're asleep. You don't even need OpenClaw for that, just a cronjob that runs claude -p in the early morning. If you give your AI enough context about yourself, you get to a point where it just independently works on things for you, and comes to you with suggestions. It doesn't need to be specifically prompted. The environment of data it can access is its own context, its own prompt. With that, it can sometimes be surprising and spooky what you wake up to, without being directly prompted."
This literally isn't even saying anything. This paragraph does not mean anything. It's not saying what its doing, whats happening or what the result is, just "something is happening".
No, you didn't save time using openclaw, you just changed to managing openclaw instead of doing your actual job.
You don't need custom scripts for most things if its actually something that matters, most tools already exist, and if you do openclaw isn't going to help you do it.
177 comments
I have experimented with a lot of hacks, like hierarchies of indexed md files, semantic DBs, embeddings, dynamic context retrieval, but none of this is really a comprehensive solution to get something that feels as intelligent as what these systems are able to do within their context windows.
I am als a touch skeptical that adjusting weights to learn context will do the trick without a transformer-like innovation in reinforcement learning.
Anyway, I‘ll keep tinkering…
I haven't seen any issues with memory so far. Using one long rolling context window, a diary and a markdown wiki folder seems sufficient to have it do stuff well. It's early days still and I might still encounter issues as I demand more, but I might just create a second or third bot and treat them as 'specialists' as I would with employees.
- I suspect that in this moment, cobbling together your own simple version of a “claw-alike” is far more likely to be productive than a “real” claw. These are still pretty complex systems! And if you don’t have good mental models of what they’re doing under the hood and why, they’re very likely to fail in surprising, infuriating, or downright dangerous ways.
For example, I have implemented my own “sleep” context compaction process and while I’m certain there are objectively better implementations of it than mine… My one is legible to me and therefore I can predict with some accuracy how my productivity tamagotchi will behave day-to-day in a way that I could not if I wasn’t involved in creating it.
(Nb I expect this is a temporary state of affairs while the quality gap between homemade and “professional” just isn’t that big)
- I do use mine as a personal assistant, and I think there is a lot of potential value in this category for people like me with ADD-style brains. For whatever reason, explaining in some detail how a task should be done is often much easier for me than just doing the task (even if, objectively, there’s equal or higher effort required for the former). It therefore doesn’t do anything I _couldn’t_ do myself. But it does do stuff I _wouldn’t_ do on my own.
Useful work includes: bug triage, matching up external user bug reports on GitHub to the internal YouTrack, fixing easy looking bugs, working on a redesign of the website. I also want to extend it to handling the quarterly accounting, which is already largely automated with AI but I still need to run the scripts myself, preparing answers to support queries, and more work on bug fixing+features. It has access to the bug tracker, internal git and CI system as if it were an employee and uses all of those quite successfully.
Meta-work has so far included: making a console so I can watch what it's doing when it wakes up, regularly organizing its own notes and home directory, improving the wakeup rhythm, and packaging up its infrastructure to a repeatable install script so I can create more of them. I work with a charity in the UK whose owner has expressed interest in an OpenClaw but I warned him off because of all the horror stories. If this experiment continues to work out I might create some more agents for people like him.
I'm not sure it's super useful for individuals. I haven't felt any great need to treat it as a personal assistant yet. ChatGPT web UI works fine for most day to day stuff in my personal life. It's very much acting like an extra employee would at a software company, not a personal secretary or anything like that.
It sounds like our experience differs because you wanted something more controlled with access to your own personal information like email, etc, whereas I gave "Axiom" (it chose its own name) its own accounts and keep it strictly separated from mine. Also, so far I haven't given it many regular repeating tasks beyond a nightly wakeup to maintain its own home directory. I can imagine that for e.g. the accounting work we'd need to do some meta-work first on a calendar integration so it doesn't forget.
An LLM context is a pretty well extended short term memory, and the trained network is a very nice comprehensive long term memory, but due to the way we currently train these networks, an LLM is just fundamentally not able to "move" these experiences to long term, like a human brain does (through sleep, among others).
Once we can teach a machine to experience something once, and remember it (preferably on a local model, because you wouldn't want a global memory to remember your information), we just cannot solve this problem.
I think this is probably the most interesting field of research right now. Actually understanding in depth how the brain learns, and figuring out a way to build a model that implements this. Because right now, with backtracking and weight adjustments, I just can't see us getting there.
I like the idea of using small local model (or several) for tackling this problem, like low rank adaptation, but with current tech, I still have to piece this together or the small local models will forget old memories.
It's not how an llm can work right now, it needs too much iterations & a much bigger dataset than what we can work with. A single time experiencing something and we can remember it. That's orders of magnitude more efficient than an LLM right now can achieve.
This field of research has been around for decades, so who's to say when there'll be a breakthrough.
In fact, LLMs are great despite our very limited understanding, and not because we had some breakthrough about the human brain.
The way an llm learns is a very interesting way of doing it, but it sure isn't what the brain is doing.
But it's indisputable.. We can get enormous results with this technique. It's just probably not the way forward for faster learning to remediate the issue of context loss.
It's like spontaneous implemention of thought experiments from yesteryear. I wonder if all this product-focused experimentation will accidentally impact philosophy of mind after all...
I think we will eventually end up with models which can be individually trained and customised on regular schedules. After that, real-time.
I’ve had a Whatsapp assistant since 2023, jailbraked as easy assistant. Only thing I kept using is transcription.
https://github.com/askrella/whatsapp-chatgpt was released 3 years ago and many have extended it for more capabilities and arguably its more performant than Openclaw as it can run in all your chat windows. But there’s still no use case.
It’s really classification and drafting.
Yes you can automate via scripting, but interacting with a process using natural language because every instance could be different and not solid enough to write a spec for, is really handy.
tl;dr: there's a place for "be liberal in what you receive and conservative in what you send", but only now have LLMs provided us with a viable way to make room for "be loosey goosey with your transput"
I mean using AI is a great way to interpret a query, determine if a helper script already exists to satisfy it, if not invoke a subagent to write a new script.
Problem with your "script" approach is how does that satisfy unknown/general queries? What if for one run you want to modify the behavior of the script?
But there's still value in people exploring new spaces they find interesting, even if they do not meet your personal definition of pareto-optimal.
> 0 legitimate use cases
My teams currently using it for:
- SDR research and drafting
- Proposal generation
- Staging ops work
- Landing page generation
- Building the company processes into an internal CRM
- Daily reporting
- Time checks
- Yesterday I put together proposal from a previous proposal and meeting notes, (40k worth)
* Telegram Health Group, created an agent to help me track sleep, recommend my supplements based on my location, remind me in the morning and evening to monitor my food. I send it images of what I eat and it keeps track of it. * Telegram Career Group, I randomly ask it to find certain kind of job posts based on my criteria. Not scheduled, only when I like to. * Telegram Coder Group, gave it access to my github account. It pulls, runs tests and merges dependabot PRs in the mornings. Tells me if there are any issues. I also ask it to look into certain bugs and open PRs while I'm on the road. * Telegran News Group, I gave it a list of youtube videos and asked it to send me news every day at 10am similar to the videos.
So far, it's a super easy assistant taking multiple personas. But it's getting a bit painful without CC subscription
https://github.com/GhostPawJS/codex
For example, for the invitations in the OP: Have Openclaw write incoming rsvps to a database, probably a flat file here, and use the db as persistent memory: OpenClaw can compose outgoing update emails based on the database. Don't even suggest to OpenClaws that it try to remember the rsvps - its job is just writing to and reading from a database, and composing emails based on the latter. ?
Does that violate the experiment, by using some tool in addition to OpenClaw?
The other common use case seems to be kicking off an automated Claude session from an email / voicetext / text / Telegram, and getting replies back. I'm emailing Claude throughout the day now, and sometimes it's useful to just forward an email to Claude and ask it to handle the task within it for me.
But I think many people criticizing the various Claws are missing out on the cronjob aspect. There's value in having your AI do work automatically while you're asleep. You don't even need OpenClaw for that, just a cronjob that runs claude -p in the early morning. If you give your AI enough context about yourself, you get to a point where it just independently works on things for you, and comes to you with suggestions. It doesn't need to be specifically prompted. The environment of data it can access is its own context, its own prompt. With that, it can sometimes be surprising and spooky what you wake up to, without being directly prompted.
Give it enough context, long term memory, and ability to explore all of that, and useful stuff emerges.
I have some issues with the article, but I agree with some of the conclusions: It's great tinkering with it if you have time to spare, but not worth using weeks of your time trying to get a perfect setup. It's just not that reliable to use up so much of your time.
I will say, it's still amongst the best tools to do a variety of tasks. Yes, each one of those could be done with just a coding agent, but I found it's less effort to get OpenClaw to do it than you writing something for each use case.
Very honest question: One of the use cases I had with OpenClaw that I'm missing now that I don't use it: I could tell it (via Telegram) to add something to my TODO list at home while I'm in the office. It would call a custom API I had set up that adds items to my TODO list.
How can I replicate this without the hassle of setting up OpenClaw? How would you do it?
(My TODO list is strictly on a home PC - no syncing with phone - by design).
(BTW, the reason I stopped using OpenClaw is boring: My QEMU SW stopped working and I haven't had time to debug).
What seems to be somewhat working for me
1. Karpathy wiki approach
2. some prompting around telling the llm what to store and not.
But it still feels brittle. I don’t think it’s just a retrieval problem. In fact I feel like the retrieval is relatively easy.
It’s the write part, getting the agent to know what it should be memorizing, and how to store it.
I’d say it’s like 85% reliable on any given task, and since I supervise it, this is good enough for me. But for something to be useful autonomously, that number needs to be several 9’s to be useful at all, and we’re no world near that yet.
I’m currently watching someone trying and failing to roll openclaw out at scale in an org and they believe in it so much it’s very difficult to convince them even with glaring evidence staring them in the face that it will not work
"The Claw."
Some of this stuff is starting to look like technologies that worked, looked promising, but were at best marginally useful, such as magnetohydrodynamic generators, tokamaks, E-beam lithography, and Ovonics.
I've removed it.
OpenClaw runs Pi in a terminal and exposes the chat thru Telegram or any chatting app. This gave the ah-ha moment to non-coders that coders had had for 6+ months prior.
>This isn’t a bug that gets fixed in the next release. It’s a fundamental constraint of how OpenClaw manages context.
Last I checked, it doesn't!
The killer usecase is letting you make whatever you want, instead of being at the mercy of what your OS/platform dictates.
Your idea of a killer idea is a whatsapp summarizer lol.
The problem is if not carefully designed it will burn through tokens like crazy.
It's a rather simple framework around an LLM, which actually was a brilliant idea for the world that didn't have it. It also came with its own wow effect, ("My agent messaged me!") so I consider some of the hype as justified.
But that's pretty much it. If you can imagine use cases that might involve emailing an LLM agent and get responses that share context with other channels and resources of yours, or having the ability to configure scheduled/event-based agent runs, you could get some use out of having an Openclaw setup somewhere.
I find the people who push insanity like "It came alive and started making money for me" and the people who label it utterly, completely useless (because it has the same shortcomings as every other LLM-based product) like Mr. "I've Seen Things. Here's the Clickbait" here, rather similar. It's actually hard to believe they know what they're talking about or that they believe what they're writing.
Sure, anything it does can be done better with specialized tooling. If you know that tooling.
The memory thing sounds like an implementation limit rather than something fundamentally unsolvable. Just experiment with different ways of organizing state until something works?
They can automate but they are not reliable. I think of them as work and process augmentation tools but this is not how most customers think in my experience.
However, here are a several legit use-case that we use internally which I can freely discuss.
There is an experimental single-server dev infrastructure we are working on that is slightly flaky. We deployed a lightweight agent in go (single 6MB binary) that connects to our customer-facing API (we have our own agentic platform) where the real agent is sitting and can be reconfigured. The agent monitors the server for various health issues. These could be anything from stalled VMs, unexpected errors etc. It is firecracker VMs that we use in very particular way and we don't know yet the scope of the system. When such situations are detected the agent automatically corrects the problems. It keeps of log what it did in a reusable space (resource type that we have) under a folder called learnings. We use these files to correct the core issues when we have the type to work on the code.
We have an AI agent called Studio Bot. It exists in Slack. It wakes up multiple times during the day. It analyses our current marketing efforts and if it finds something useful, it creates the graphics and posts to be sent out to several of our social media channels. A member of staff reviews these suggestions. Most of the time they need to follow up with subsequent request to change things and finally push the changes to buffer. I also use the agent to generate branded cover images for linkedin, x and reddit articles in various aspect ratios. It is a very useful tool that produces graphics with our brand colours and aesthetics but it is not perfect.
We have a customer support agent that monitors how well we handle support request in zendesk. It does not automatically engage with customers. What it does is to supervise the backlog of support tickets and chase the team when we fall behind, which happens.
We have quite a few more scattered in various places. Some of them are even public.
In my mind, the trick is to think of AI agents as augmentation tools. In other words, instead of asking how can I take myself out of the equation, the better question is how can I improve the situation. Sometimes just providing more contextually relevant information is more than enough. Sometimes, you need a simple helper that own a certain part of the business.
I hope this helps.
First of all is not an LLM, you're beholden to an api or local llm limitations. Second of all it's always calendars, email replies, summarizing.
You do not need an LLM for that, and an LLM doesn't make it easier either. It sounds like executive cosplay, not productivity. Everything I see people talking about that's actually productive, it's doing probabilistically when deterministic tools already exist and have for in some cases over 20 years.
You don't need an LLM to put a meeting on a calendar, that's literally two taps with your phone or a single click in gmail. Most email services already have suggestions already built in. Emails have been summarized for 10 years at this point. If you're so busy you need this stuff automated, you probably have an assistant, or you're important enough that actually using general intelligence is critical to being successful at all.
The idea of getting an LLM email response sounds great for someone who has never worked a job in their life.
This comment section is full of llm writen responses too, to the point where its absurd. Noticing how most of them just talk in circles like "But I think many people criticizing the various Claws are missing out on the cronjob aspect. There's value in having your AI do work automatically while you're asleep. You don't even need OpenClaw for that, just a cronjob that runs claude -p in the early morning. If you give your AI enough context about yourself, you get to a point where it just independently works on things for you, and comes to you with suggestions. It doesn't need to be specifically prompted. The environment of data it can access is its own context, its own prompt. With that, it can sometimes be surprising and spooky what you wake up to, without being directly prompted."
This literally isn't even saying anything. This paragraph does not mean anything. It's not saying what its doing, whats happening or what the result is, just "something is happening".
No, you didn't save time using openclaw, you just changed to managing openclaw instead of doing your actual job.
You don't need custom scripts for most things if its actually something that matters, most tools already exist, and if you do openclaw isn't going to help you do it.