OpenClaw 4.29: New NVIDIA + Memory Updates

OpenGLore 4.29 just dropped and this one reworks how agents behave in group chats. Your agent should no longer blurt out every thought the second it has one. It's supposed to think, use tools, and then deliberately send the message when it's ready. On top of that, there's a new follow-up system for commitments, basically where your agent picks up on things you mention and checks back with you later to hold you accountable.

It builds a people wiki from your conversations. NVIDIA joined as a built-in provider and there are startup and channel improvements across the board. Let me walk you through what's new. Now before I get into the features, I want to be upfront about something.

OpenGLore has been rough the last few releases. A lot of people in the community have been reporting bugs, rollback issues, broken local model setups. So whilst I'm going to tell you what's in this update, I'm not going to promise everything just works perfectly out of the box. Test things yourself and if something breaks, you can always restore a previous version.

So let's start with group chats. This is one of the biggest changes in OpenGLore 4.29. Before this update, when your agent was in a Discord server or a Telegram group or a Slack channel, it would automatically post its reply the moment it finished processing, even if it was still running tools, even if the reply wasn't polished yet. The final assistant text just got dumped into the chat.

The new behavior changes that. Your agent is supposed to control when it speaks, it thinks, it runs tools, it checks things and then it deliberately sends a message using the message tool. So what shows up in the group should be intentional and by default, group replies are now private. Your agent shouldn't automatically post in the channel anymore.

It waits until it has something worth saying and it says it on purpose. If you want the old auto-posting behavior back, there's a setting called messages.groupchat.visiblereplies that you can set to automatic. There's also a new global setting, messages visible replies, that lets you control this across all conversations, not just groups. If this works as described, it's a big deal for anyone running agents in busy group chats, for agencies running client groups, for communities with AI systems, for team channels.

Your agent should behave more like a thoughtful participant instead of just an agent that can't stop talking. But again, test it with your setup before you rely on it. Now here's something new, follow-up commitments. This is opt-in, it's not on by default.

Turn it on and your agent starts picking up on things you mentioned in conversation that need a follow-up later. Not because you told it to set a reminder, but because it understood your context. Say you're chatting with your agent about a client proposal and you mention, I need to send the final version by Friday. Your agent picks it up, it creates a commitment internally and when Friday comes around, it checks in with you through the heartbeat system and says something like, for example, hey, did you send that proposal?

You control how many commitments it can create per day with commitments.maxPerDay. You can scope it per agent or per chat and the extraction happens in the background so it shouldn't slow down your conversation. The due times get clamped to heartbeat intervals so you won't get pinged immediately after mentioning something. For anyone running a business, think about how many things fall through the cracks in conversation.

A client mentions a deadline, you tell yourself you'll check on a task later. If this feature works well, your agent catches those and brings them back at the right time. I'm curious to see how it performs in real usage. Now let's talk about memory because it got a big upgrade too.

Your agent now builds what they're calling a people wiki. As you have conversations, your agent starts tracking the people you mention. It builds cards for each person with aliases, relationships, and context about how you know them. It tracks where information came from, which conversation, which date, what kind of evidence.

So if you mention your client Sarah in three different conversations over two weeks, your agent should connect the dots. It knows Sarah is a client, it knows what project she's on, it knows the last time you discussed her, and when you ask about Sarah later it pulls from that wiki instead of searching through old conversations. There are search modes for person lookup, question routing, source evidence, and raw claims. There's a privacy and providence report so you can see exactly what your agent knows and where it learned it, and there's a relationship graph that maps connections between people.

The recall system also got more resilient at least in theory. Before if memory search took too long it would fall and fail and your agent would no longer reply with context. If there's a timeout now it's supposed to return whatever it found so far. Partial results instead of nothing, and there's per conversation filtering for active memory so you can scope recall to specific chats.

If you want help setting up memory systems, setting your agent, helping you automate with agents like OpenClaw and Hermes, come join us in the AI Profit Boarding. We run four weekly coaching calls every week where members can bring their OpenClaw setups and we can work through how to configure it. You can ask questions in real time, you can share your screen. We also have 30-day roadmaps and a full six-hour course on OpenClaw, a full two-hour course on Hermes, and daily tutorials on new features dropping too.

Plus you can connect with me personally and there's a map so you can connect with people using AI agents in your city. Link in the comments description or go to the AIProfitBoarding.com to get access. NVIDIA is now a built-in provider. If you're using NVIDIA's hosted AI models, Nemetron or anything else they host through their API, you can now set them up inside OpenClaw with your NVIDIA API key.

There's proper onboarding, a static catalog so you can browse models and model names work with their provider prefix intact. Model catalogs in general move toward manifest-first metadata. That means your model list should load quicker because the data comes from plugin manifests instead of being rebuilt on every startup which means less scanning, faster results. There's a new message steering system too.

Before if you sent a message whilst your agent was already working on something, it would either get dropped or create a duplicate run. Now your follow-up gets injected into the active point run at the next save point. Your agent sees it and adjusts it. No duplicate work, no dropped messages.

The default is steer mode with a 500 millisecond debounce. There's also a queue mode if you want the old behavior. The exec and security stuff got tighter too so restrictive tool profiles now stay restrictive. Before certain tool sections could accidentally widen permissions even when you set messaging or minimal mode.

That shouldn't happen anymore. Owner checks for pairing and device tokens are stricter and there are setup warnings that flag risky configs early. For channels there are improvements across the board. Slack should handle black kit limits better now.

Oversized button values and interactive elements get trimmed instead of creating and crashing the whole message. Telegram got proxy and webhook resilience improvements. Discord got rate handling limits for Cloudflare 49s during startup. WhatsApp should confirm delivery before marking messages as sent.

The keyword is should. Given recent releases I would just test all of this carefully before trusting it in actual production. The gateway is supposed to start faster now too. Channels should come online before model warm-up finishes and there's a new readiness endpoint that shows event loop health so you can see what's happening during startup.

For windows users specifically plug-in loading now uses a faster native path. The claim is startup dropped from about 39 seconds to about two seconds on typical setups. That will be huge if it holds up. Now here's the important part.

How to update. Before you do anything create a backup. Run open claw backup create. This saves your config, your sessions, your memory files.

If something goes wrong you can roll back later. To update run open core update. It will download the new version, sync your plugins and restart the gateway. If you're on npm you can also run npm install dash g open core at latest.

But here's my honest advice. Given how the last few releases have gone 4.24 through 4.27 they all had issues that people sometimes needed to roll back from. Don't update on a machine that's running something critical for you right now. Update on a test setup first.

Make sure your channels connect. Make sure your local models work. Make sure your agents respond quickly and then if you're happy with it then update your main setup. If something does break and you need to roll back you can run open core update dash dash channel stable and pick an earlier version or if you saved your backup restore from that.

There were 198 contributors that shipped code for this release. The community keeps growing and the direction is right. Group chats aren't chaotic. Memory that tracks people.

Agents that follow up on their own. These are the features that make AI agents actually useful for your business. But features only matter if they work and I've seen too many people in the community hit problems after updating lately. So be careful.

Backup first. Test before you trust it and if things work well for you great. You've got some powerful new tools to play with. The people who are building with open core now, learning the setups, configuring the agents, testing each release they're the ones who have a real advantage when all of this stabilizes.

Every release gets closer to what it should be and when it's there the people who already know how to use it won't need to play catch up. Come join us in the AI Profit Boarding. We've got tutorials on the new updates, how to set it up, how to use OpenCore in your business. We've got full courses on how to run this.

Four coaching calls a week. Loads of community members using AI agents like OpenCore and Hermes. Daily tutorials and a prompt library for these sort of workflows. There's 2,900 members in there as well and a member map so you can connect with other OpenCore users near you.

Link in the comments description or go to the AIProfitBoarding.com to get access. I'll see you in the next one. Thanks for watching.