OpenClaw v2026.4.29 landed on April 30, 2026 at 4:01 p.m. America/Chicago time, which was April 30, 2026 at 21:01 UTC. The release is broad, but the useful summary is simple: your follow-up messages are less likely to get lost while a run is active, memory now understands people better than before, and the gateway keeps shedding the sort of startup and channel bugs that waste an afternoon.

This one also hides an important config change inside the long changelog. If you run a restrictive tool profile such as messaging or minimal, configured tools.exec or tools.fs sections no longer widen access by accident. That is a good safety move. It can also surprise you after an update if you were relying on the old behavior.

Active-run follow-ups finally look more intentional

The biggest day-to-day behavior change is steering. OpenClaw now defaults active-run queueing to steer, not the old one-message-at-a-time behavior. If you send a follow-up while the assistant is still working, the system is much more willing to treat that as real guidance for the current run instead of awkward backlog.

That matters if you use OpenClaw like a real coworker. You change your mind mid-task. You add one more file. You clarify what "latest" means. This release moves that from "maybe it catches up" to something you should actually test and start trusting.

There is also new visible-reply enforcement and opt-in follow-up commitments. The first is for operators who want cleaner guarantees about where replies show up. The second lets OpenClaw schedule lightweight check-ins without blasting you immediately after a heartbeat. Those are not flashy features. They are the kind you notice after a few days.

Memory stops treating people like loose scraps of text

Memory grows into a people-aware wiki. That means aliases, person cards, relationship graphs, provenance views, and better ways to search for who someone is instead of only matching raw snippets. If you keep long-running chats, shared projects, or recurring contacts in the loop, this is a real upgrade.

I like the smaller memory changes too. Active Memory can now be scoped to specific chats, and timed-out recall can return a bounded partial result instead of dropping useful context on the floor. You should feel that as fewer "I almost remembered it" moments.

There is a lot of operator-grade work in here

NVIDIA joins the provider list. Startup diagnostics get a proper timeline. Model and auth lookups keep moving onto faster manifest-backed paths. Gateway restarts and stale-session recovery got more attention. The packaged-plugin runtime-deps story keeps getting repaired in places where it used to fail at exactly the wrong time.

If you only care about one happy-path chat, that reads like plumbing. If you run OpenClaw every day across channels, plugins, or multiple providers, it is the difference between a clean restart and another round of "why did this break only on this machine?"

Channel reliability still gets the boring fixes it needed

Telegram picked up more durability work around polling, webhooks, quote replies, proxy behavior, and outbound sends. Slack got another round of fixes for Block Kit size limits, command menus, interactive replies, and approval cards. Discord startup and rate-limit handling got harder to wedge. WhatsApp, Signal, Feishu, Matrix, and Teams all picked up edge-case repairs too.

That is a lot of surface area, but the practical takeaway is easy: if your assistant lives inside chat apps, this release is worth it for the reliability work alone.

What I would test right after updating

1. Interrupt one active run on purpose. Send a task, then send a follow-up while it is still running. Make sure the second message steers the work the way you expect.
2. Ask memory about a real person. Pick a teammate, customer, or recurring contact and see whether the answer comes back cleaner than a bag of unrelated snippets.
3. Restart the gateway once while you are watching. This release spends a lot of energy on startup, stale-session recovery, and dependency repair. It is worth checking on your host before you need it under pressure.
4. Test your main chat surface. Send one normal reply and one reply with media or structure if you rely on Slack, Telegram, Discord, or WhatsApp.
5. Audit restrictive tool profiles. If you depend on messaging or minimal plus extra tool sections, add explicit alsoAllow entries where needed and confirm the startup warning is gone.
6. If you use Active Memory, check chat scoping. The new allow and deny chat filters are useful, but only if they match the conversations you meant to include.

Should you update now?

Yes if you run OpenClaw heavily in chat, rely on long-lived memory, or ever send follow-up corrections while the assistant is still mid-run.

Also yes if you operate a busier install with plugins, local models, several channels, or provider mix-and-match. A lot of this release is about keeping the system from going sideways in the messy cases.

The cleanest way to describe v2026.4.29 is that it helps OpenClaw behave more like a teammate and less like a queue you have to babysit.

---

Full changelog: View v2026.4.29 on GitHub

Generated release page: Read the structured release notes

- Fred