OpenClaw 4.15 Update: LanceDB Cloud Memory and OAuth Health Card
Prefer to watch? Watch on YouTube · Subscribe for new builds · Read the transcript
In this video I go through the OpenClaw 4.15 release and explain which changes actually matter for your agent setup. It's for anyone already running OpenClaw who wants to know whether to upgrade, and for newer builders who want to set up memory the right way from the start. I recorded this in April 2026. Tools like this change fast, so check the current docs before you configure anything.
What I cover in this video
- 0:00: The changelog headlines: Opus 4.7 as the Anthropic default, Google text-to-speech, the Control UI overhaul, and LanceDB cloud storage
- 3:53: How OpenClaw memory worked before 4.15, and why local-only memory is a problem
- 5:25: What LanceDB cloud memory changes, including shared memory across machines
- 6:55: The OAuth health card and the silent failure it catches
- 8:27: The time a broken Claude OAuth token quietly hurt my content pipeline for days
- 9:13: Slim mode for local models, plus GitHub Copilot embeddings
- 11:29: Telegram topic persistence and the doctor/systemd security fix for Linux
- 13:01: What 4.15 means for your setup, and how to update
The headline changes in 4.15
At the top of the changelog, Anthropic selections now default to Opus 4.7. That includes the Claude CLI defaults and bundled image understanding. Opus 4.7 can get expensive. But if you're selling agents as a service or building a product on top of OpenClaw, you want the best output, and right now that's Opus 4.7.
The release also adds Google text-to-speech. That's a small upgrade if you're already on a Mac mini or using Apple's voices. The Control UI got a big overhaul too, including a model auth status card that shows OAuth token health and provider rate limit pressure at a glance.
The OpenClaw team has been shipping a lot. We went from 4.10 to 4.15 in about a week. Most of those releases were bug fixes and performance tweaks. 4.15 is different because two of its features change how you'll actually use the tool.
LanceDB cloud memory: your agent's memory isn't tied to one machine anymore
As your OpenClaw agent works, it builds up memory about you: your preferences, your business, your workflows, what you like and what you don't. Before 4.15, all of that lived locally on the machine where OpenClaw was installed. If you ran OpenClaw on a home server and on a laptop, the two installs didn't share any memory. Each one was its own isolated bubble.
That becomes a bigger problem the longer you use OpenClaw, because the tool gets more valuable as it builds up context. You spend weeks feeding it your business rules, your voice, and your priorities until it feels dialed in. Then you get a new machine, wipe a server, or do a clean reinstall, and all of that context is gone. I've lost everything before. The only thing that saved me was that I already had memory backed up in the cloud.
With 4.15, LanceDB supports cloud storage, so durable memory indexes can run on remote object storage instead of only on local disk. In practice, that means:
- Two machines running OpenClaw can point to the same memory store and share the same context and history.
- When you reinstall or move to new hardware, your memory comes with you.
- A new machine you add to your setup inherits everything from day one.
If you're just getting started, set this up on day one. Don't wait until you have months of valuable memory sitting in a local folder that's one disk failure away from disappearing. I also think you should have more than one backup. Personally I use supermemory.ai, Neon for unstructured data, and a push to GitHub every time the heartbeat goes off (every 2 hours).
The OAuth health card: catching the failure you can't see
For most everyday users, this is the update with the bigger impact. A lot of people, especially those who started with a Claude subscription instead of raw API access, have an OAuth token running in the background that handles their connection to Claude. When that token expires or becomes invalid, the agent doesn't crash or throw an error. It just quietly gets worse: slower responses, shorter outputs, and more generic answers.
Nothing tells you to investigate. Tasks still complete, but the agent feels less sharp. Unless you dig through your logs, you'll probably blame the model, your prompts, or your whole setup. The actual fix is to re-authenticate, which takes about 30 seconds.
This happened to me in late March. My content pipeline started producing softer scripts and weak research. I spent a couple of days tweaking prompts and reworking my setup before I found the cause: Claude OAuth had been failing silently, and the system had fallen back to a cheaper model without telling me.
The new card puts your token status right on the Control UI dashboard. Green means you're good. If a token is expiring or expired, you get an attention callout as soon as you open the dashboard. You don't have to run commands or read logs.
Other updates worth knowing
- Slim mode for local models: If you run Ollama or another local inference setup on lower-spec hardware, the standard install's overhead competes with the model for resources. Slim mode strips the runtime down so your CPU and RAM go to the model. It's worth turning on for home servers and low-spec VPS setups.
- GitHub Copilot embeddings: If you already pay for Copilot, you can use it as your embedding provider instead of setting up a separate embedding service.
- Telegram topic persistence: Before 4.15, OpenClaw forgot which Telegram topic names mapped to which channel IDs every time it restarted. Those mappings now persist.
- Doctor/systemd security fix: On Linux, the repair command had a bug that re-embedded secrets into systemd unit files. That's patched. If you run OpenClaw on Linux and have used the doctor or repair commands, that alone is a good reason to update.
To update, run this in your terminal, or just ask your agent to update itself:
openclaw update
Key takeaways
- LanceDB cloud memory lets your agent's memory live on remote storage, so it survives reinstalls and can be shared across machines.
- Personal AI agents get more valuable as they build up context, and that value disappears if it's tied to a single disk.
- An expired Claude OAuth token can quietly downgrade your output. The new health card makes that visible, and re-authenticating fixes it.
- Slim mode, Copilot embeddings, and Telegram topic persistence are useful quality-of-life changes for specific setups.
- Linux users who have used the doctor or repair commands should update for the security fix.
FAQ
What is LanceDB cloud memory in OpenClaw 4.15?
It lets OpenClaw's durable memory indexes run on remote object storage instead of only on local disk. Multiple machines can share the same memory store, and your memory survives reinstalls or moving to new hardware.
Why does my OpenClaw agent feel worse even though nothing is broken?
Your Claude OAuth token may have expired or become invalid. When that happens the agent doesn't error out, it just quietly degrades and can fall back to a cheaper model. Re-authenticating fixes it, and the 4.15 OAuth health card shows the token status on the Control UI dashboard.
How do I update to OpenClaw 4.15?
Run openclaw update in your terminal, or ask your agent to update itself.
What is slim mode in OpenClaw?
Slim mode strips the OpenClaw runtime down to what people running local models actually need, so more of your CPU and RAM goes to the model. It helps most on home servers and low-spec VPS setups.
Chapters
Transcript
Read the full transcript
0:00Openclaw 4.15 just came out and we have Opus 4.7 native support Lance DB for local and cloud storage as well as better security features. And in this video we're going to be breaking down all of them on why you should be upgrading if you are on an older version. Okay, so this just came out last night. Here's the change log and the first one at the top of the list is Anthropic models. So Anthropic selections of Opus 4.7 now by default. You can see the cloud CLI defaults and bundled image understanding to cloud Opus 4.7. Yes, it can be expensive, but if you're running a significant setup and you're running, you know, agents as a service and selling your product or building on top of Openclaw to sell a product, you're going to want the best output. So Opus
0:474.7 is the front runner here as well as Google text-to-speech. Not really a huge upgrade for people, you know, that are already using Apple or Mac minis, but that is a nice upgrade there as well as the control UI got a big overhaul. They added model off status card showing OAuth token health and provider rate limit pressure at a glance with attention callouts when the OAuth tokens are expiring or expired. So just better back-end support as well as the memory. The memory is always a constant issue when it comes to Openclaw. So you can see Lance DB it added cloud storage support to So durable memory indexes can run on remote object storage instead of local disk only. So those are the features that kind of highlight this
1:33whole release, but we're going to be deep diving on why it actually matters and why you should upgrade. I think two of those features genuinely move the needle. The one with memory, I think that's a huge thing. It's a constant problem. You know, if you've been using Openclaw for just any length of time, you already know that memory is the number one thing people go crazy over. The other one is a fix for a problem that has been quietly making your agents basically perform worse without giving you any signal that something's wrong. So for some quick context, the Openclaw team has been on an absolute tear this month. I mean, there has to be at least 10 updates in the last 15 days. So we went through We went from 4.10 to 4.15
2:20within a week, which is pretty crazy. I love to see it. Real quick, before we keep going, if you're watching this and you want to actually build with some of these tools, not just watch videos about them, you're going to want to check out our community down below, Shipping School. We have a full Claude code course, a full Openclaw course, and four live bootcamps every single week where we actually help you get set up from scratch. Like actually set this thing up, not just watch a tutorial and figure it out by yourself. And we also provide one-on-one coaching, so you could book a call with me, we could share screens, and I can help you get Claude code or Openclaw running on your machine. That's it, no fluff. I built this community because watching YouTube's only gets you so far. We launched it just 3 days ago and we have over 55 members. You need
3:08people around you who are actually building, people who hold you accountable, and coaches who can help you when you get stuck. I'll put the link in the description down below. Get in now before the price goes up. And most of those releases in between were small quality of life things, you know, nothing huge, but like bug fixes, performance improvements, you know, those kind of things, but 4.15 has a headline feature that is in a completely different tier. That feature is the Lance DB cloud memory. And it basically changes how the whole memory layer works in a pretty fundamental way. So this, in plain terms, I'm going to just break it down because the technical explanation makes it sound basically way more complicated than it needs to be for most of us users. So when your Openclaw agent
3:53runs and does work, it builds up memory, right? Context is generated about you, your preferences, your business, your workflows, what you like, what you want it to remember, what you don't like. Right now, in every version before 4.15, all of that memory lives locally on a specific machine where Openclaw is installed. So if you run Openclaw on a home server and you also want it running on your laptop, those two installs do not share memory at all. Machine A has no idea what happened on machine B. Every install is its own isolated bubble. And this becomes a huge problem the longer you use Openclaw because the value of the tool compounds with context. You spend 2 weeks training your agent, feeding it your business rules,
4:39your preferences, the decisions you've made, the workflows that actually matter, and then your agent starts to feel genuinely dialed in. You think you're feeling pretty good, you know, but, you know, things change. It can know your voice. Like I have a voice profile and you feel like you're making progress. It can know your priorities. It knows what our priorities are. It anticipates what you need even. Then you get a new machine or your server needs to get wiped clean or you just do a clean reinstall. And all of that built-up context is gone. And I've experienced that before when I lost everything and had to restart over, but I had memory in the cloud. So you have to start from scratch otherwise. And that is months of compounding value wiped out by a hardware event. We don't want that. So what they have is they basically have that cloud memory built
5:25in with this newest update. With Lance DB, cloud memory solves all of this. What it means in practice is that your vector storage where your agent's memory lives can now be cloud hosted instead of sitting on a local disk. So you can have two machines running Openclaw and both of them point to the same memory store. Your server and your laptop know the same things. They share the same context, the same history, the same accumulated understanding of your business. When you reinstall or move to new hardware, your memory follows you. When you add a new machine to your setup, it inherits everything from day one. For people running a multi-machine setup, this is a massive unlock. For anyone who has ever migrated to new hardware and felt that gut punch of
6:10realizing months of build-up context just evaporated overnight, you know exactly what this fixes. And for people just getting started with Openclaw, this is something you want to configure on day one rather than discover the hard way later when you have a bunch of valuable memory sitting in a local folder that is one disk failure away from disappearing. I mean, I think that would be very demoralizing. You got to figure out how am I going to back up all of this progress that I've made? I have supermemory.ai. I have Neon for unstructured data. I push to GitHub when the heartbeat goes off every 2 hours. I have multiple ways to back up all of this memory. And I think you should be thinking the same way, too. I want to talk about the second big thing in this update because I think for most everyday users, this is
6:55one that actually more people that you would care about, right? I think this is more impactful. It is the OAuth health card in the control UI. And let me tell you the problem with this solves, right? A lot of Openclaw users, especially people who started out using Claude subscription rather than buying the raw API access or going through the tokens, have an OAuth token running in the background that handles their connection to Claude. That token works great until it doesn't. It expires or some something upstream changes or it gets invalid for some reason. And when that happens, your agent does not crash. It does not throw an error. It does not tell you anything wrong. It just quietly starts degrading. Slower responses, shorter outputs, less nuance, more generic answers where you
7:40used to get the sharp, precise ones. The insidious thing about this is the failure mode is there, but it's nothing to trigger the investigation. Your agent is still running. Tasks are still completing. You just gradually get this sense that something's off. And I know we've been there before. You know, the agent just doesn't feel as sharp anymore as it used to. And that The outputs have gotten pretty flat. And unless you go digging into your logs or think to run a diagnostic, you will never connect the symptom to the cause. You will assume the model is having a bad stretch or that the Claude servers are down or that your prompts have gotten stale or that you need to rebuild your entire setup. Don't do that. Just The fix is literally just reauthenticating, which takes 30 seconds. I ran into this myself back in late March. My whole content pipeline
8:27started feeling off. Scripts were coming out softer, the research outputs were just terrible. The overall quality was just bad. I spent a couple days tweaking the prompts, reworking my setup before I finally figured out the Claude OAuth had been silently failing. And the system had quietly fallen back to a cheaper model without surfacing that to me. That one broken token was affecting my production output for days before I caught it. The OAuth health check card puts your token status directly on the dashboard. So green means good to go. If something has expired or broken, you see it the moment you open your control UI. You don't need to run commands, you don't need to parse logs, and you don't need any technical knowledge to catch it. It's just there. For anyone running a pipeline where agent quality matters, this is the one
9:13of those things where you need to You should have been there from day one, honestly. I mean, come on. There are a few other things in 4.15 worth going through. Slim mode for local models is a good one. If you run a llama or any other local inference setup, the standard Openclaw install carries a bunch of resource overhead that makes sense when you are pointed at a powerful remote model and want the full-featured environment. But when you are running a smaller local model on a lower-spec hardware, that overhead competes with the model itself for resources. Slim mode strips the runtime down to what local model users actually need. So your CPU and RAM budget goes to the model rather than infrastructure. So if you run Openclaw on a home server or a low-spec VPS or whatever, or you're
9:59experimenting with local models in general and you have been feeling that resource squeeze, slim mode is worth turning on. They also added GitHub Copilot embedding support to this release. So, if you're already paying for a Copilot subscription, you can now use it as your embedding provider directly instead of spinning up a separate embedding service or paying for another API. It is a consolidation move for anyone living in the GitHub ecosystem. One less billing line to manage, one less service to configure. It's pretty straightforward. Telegram topic persistence is also in here and this is one of those things that sound small, but it has been causing real headaches for a lot of people. If you're struggling to keep up with content, what I'm about to save you about 40 days
10:44worth of work. I built something called Content Machine. [music] It's 10 AI agents that run on the Open Claw orchestration and they handle everything. Scripts, [music] thumbnails, X posts, blogs, outreach, clips, newsletters, all of it. So, I went from 1,000 subscribers [music] to 4,000 subscribers on YouTube in 7 days using this exact system. Every single morning I wake up [music] and the content's already done. I spend maybe 15, 20 minutes reviewing and approving them and I move on with my day. It works for any niche, [music] fitness, finance, real estate, marketing, whatever you are building and it is 100% completely customizable to your [music] use case. So, you get the mission control dashboard, all of the
11:29cron jobs, everything I've built over the last 40 days helping me gain more and more people to subscribe and join the community. So, you plug in your own thing and it molds it to you, it learns how you talk and it writes so it doesn't sound like AI slop. $97 one-time, it's not a subscription. I'll put the link down below and you'll thank me later. Before 4.15, every time Open Claw restarted, it would forget which Telegram topic names map to which channel IDs internally. That can cause a lot of headache. So, there was this annoying relearning cycle every time the service came back up. If you run a live school community like me or a customer support channel through Telegram and you have been dealing with that reset behavior after restarts, that is now
12:16fixed. Topic mappings persist, it's all done. And then there is a doctor in systemd fix that matters specifically for people on Linux servers. The repair command had a bug. It was accidentally re-embedding secrets into systemd unit files during the repair process. That is a security concern and it is patched now in 4.15. If you run Open Claw on Linux and you have ever used the doctor or repair commands, updating today is the right call for just that reason alone. So, stepping back and thinking about what 4.15 actually means for people building with this stuff, here's just basically how I think about it in real simple terms. The cloud memory feature is the strategic one. The whole point of personal AI agents is that it gets
13:01smarter over time because it accumulates context. It learns how you think, what your business needs, what decisions you have made and why, but all that compounding value is worthless if it is anchored into a single machine. Portable persistent memory is what makes the agent relationship durable. Your investment in building up your agent context now follows you instead of being held to your hardware situation. That is a fundamentally more resilient setup and for anyone who has ever lost their context to a machine change, you know exactly why this matters. The OAuth health card is a practical one. The worst failure mode for any production setup is the invisible failure. How can we detect these? That is all fixed now. The thing that degrades your output without telling you anything is wrong, visibility is the fix. When you can see
13:47what is happening, you can respond to it. When failures hide in the background, they erode your trust in the whole system and eat up time you should be spending on actual work. If you're running Open Claw right now, go update to 4.15. The command is just open claw update in your terminal or just ask your agent to update itself. It handles everything. And if you're just getting into this, trying to figure out how to build your first real agent setup, we have a whole community based on this called Shipping School where we actually help people build exactly these kinds of setups together, sharing what works, what breaks and how we put it all together without needing a technical background. I'll put the link down below and we're actually hiring in the group, too. If you are a teacher or you've built something amazing with Open Claw, if you go to shippingschool.com, we have a careers tab. We pay $100 to $1,000 an
14:34hour depending on experience. So, you can check that out like I said, link down below. And also, if you haven't subscribed to the channel already, please do cuz we release videos every day covering AI news and Open Claw updates. We'll see you in the next one. Have a blessed day.
Ready to start building with AI?
Join the Shipping Skool community free. Paid members get live bootcamp calls every week and weekly-updated courses.
Join Shipping SkoolKeep reading
OpenClaw 4.22 Update: Streaming Voice Agents, Grok, and /models add
OpenClaw 4.22 lets you build real-time voice agents with five streaming providers (OpenAI, xAI, Deepgram, ElevenLabs, Mistral). It also adds Grok image generation and six TTS voices, a TUI mode that runs without the gateway, /models add for adding models from chat, and faster plugin loading.
OpenClaw 4.12 Update: Active Memory, LM Studio, and Security Fixes
OpenClaw 4.12 adds an optional Active Memory plugin. Right before your agent replies, a memory sub-agent pulls in your relevant preferences and past details, so you don't have to keep saying "remember this." The release also adds built-in LM Studio support, a new exec policy CLI, three security patches, and fixes to dreaming.
OpenClaw 4.9 Security Update, Why You Need to Upgrade Right Now
OpenClaw 4.9 is not a cosmetic update. It fixes real security issues, upgrades agent memory, and gives builders better model testing before deployment.
