Two releases in 48 hours. That’s the pace OpenClaw is setting right now — and if you’re running an AI assistant for your business, you need to know what just changed. On April 11, 2026, OpenClaw shipped version 2026.4.10 — a 17-feature, 20+ fix release that rewrites how the platform handles memory, coding agents, and voice. Then, just two days later on April 13, version 2026.4.11-pre followed with 9 more enhancements and 12 stability fixes focused on rounding out those new capabilities. Together, these two releases represent the most significant upgrade cycle since OpenClaw 4.8. Here’s what’s actually new, what it means for your operations, and why you should be planning your upgrade now.

What’s New in OpenClaw 4.10

1. Native Codex Integration — Your Coding Agent Just Got Serious

Until now, if you wanted to use OpenAI’s Codex models through OpenClaw, they shared the same authentication and threading path as regular GPT models. That caused conflicts — wrong auth flows, broken compaction, mixed contexts. 4.10 fixes this by making Codex a first-class provider. When you configure codex/gpt-* models, they get their own:

**OAuth authentication** (separate from your OpenAI API keys)
**Native thread management** for long coding sessions
**Automatic context compaction** so your agent doesn’t lose track after hours of work
**Independent model discovery** and configuration

Your existing openai/gpt-* setup is completely untouched. This is a parallel path, not a replacement. Why it matters for business: If you use OpenClaw to run coding agents (Claude Code, Codex, Pi), this removes a major pain point. Long sessions stay coherent. Auth doesn’t break mid-task. Your development team’s AI workflows become more reliable overnight.

2. Active Memory Plugin — Your AI Remembers Without Being Asked

This is the biggest user-facing change in 4.10, and it fundamentally changes how OpenClaw *feels* to use. Previously, OpenClaw’s memory system was reactive. You had to explicitly say “remember this” or the agent had to manually search memory files. Useful context from everyday conversations was routinely lost. Active Memory adds an automatic memory sub-agent that runs before every reply. It pulls relevant preferences, context, and historical details — without any manual trigger. You mention you’re heading to Tokyo next week in a casual message, and the agent just… remembers. Next time you ask about flights, it already has the context. Three modes let you control the tradeoff: |——|————–|————|———|

Mode	Context Scope	Token Cost	Latency
`message`	Current message only	Minimal	~1-2s
`recent`	Recent conversation history	Moderate	~2-3s
`full`	Entire memory store	Higher	~3-5s

Why it matters for business: This is the difference between an AI assistant that *responds* and one that *understands*. For teams using OpenClaw across client accounts, project management, or daily operations, Active Memory means less repeated context-setting and more productive conversations from the start. Recommendation: Start with message mode. Monitor token consumption for a week. Upgrade to recent if the latency is acceptable for your use case.

3. Local MLX Voice — Talk Mode Without the Cloud

If you’re on a Mac with Apple Silicon (M1 or later), 4.10 brings local voice inference to OpenClaw’s Talk Mode. No cloud TTS dependency. No latency from sending audio to ElevenLabs and waiting for a response. This is marked experimental, but the implications are clear:

**Privacy:** Voice conversations never leave your hardware
**Reliability:** Works offline, no network dependency
**Speed:** Local inference on Apple Silicon is fast enough for real-time conversation
**Fallback:** Automatically switches to system voice if the MLX model can’t handle something

Why it matters for business: For any operation that handles sensitive client information — legal, financial, medical — local voice processing eliminates a compliance concern entirely.

4. Security Hardening — SSRF Protections and Browser Safety

4.10 includes comprehensive browser and sandbox hardening:

**Strict SSRF defaults** to prevent Server-Side Request Forgery attacks
**Hostname whitelisting** for controlled network access
**CDP (Chrome DevTools Protocol) origin scope enforcement** in Docker
**Subframe and interaction-driven redirect protections**

These aren’t theoretical improvements. SSRF is one of the most common attack vectors for self-hosted AI platforms, and these changes close real vulnerabilities. Why it matters for business: If you’re running OpenClaw on a VPS with customer data, these security improvements aren’t optional. They’re the reason you can tell your clients their data is safe.

Additional 4.10 Features Worth Noting

**Seedance 2.0 video generation** through the fal provider — duration, resolution, and audio controls
**Message actions:** Pin, unpin, read, react, and list reactions across channels
**Exec policy management:** `openclaw exec-policy` command for quick security preset switching
**Command discovery API:** Remote clients can now discover available commands at runtime
**Private network model access:** Connect to self-hosted vLLM or Ollama instances with proper per-provider configuration
**Strict-agentic Pi execution:** GPT-5 models now work through filler turns until hitting genuine blockers

What’s New in OpenClaw 4.11-pre

If 4.10 was about new capabilities, 4.11-pre is about making those capabilities reliable and usable.

1. ChatGPT Import Integration in Dreaming

OpenClaw’s Dreaming system (its sleep-time memory consolidation) can now:

**Import ChatGPT conversation histories** directly into the memory pipeline
**Display imported source chats** in the UI alongside compiled wiki pages
**Provide full traceability** — see which original conversation a preference or memory was extracted from

This turns Dreaming from a black-box memory process into an auditable knowledge workbench. You can trace any piece of stored memory back to its original source conversation. Why it matters for business: If your team has years of ChatGPT conversation history with client context, project decisions, and institutional knowledge — that data can now flow into OpenClaw’s memory system instead of being locked away.

2. WebChat Rich Media Rendering

WebChat graduates from a plain text window to a structured interactive console:

**Voice replies** render as proper audio bubbles
**Media embeds** (video, images) display inline with configurable embed whitelists
**Reply directives** appear as interactive chat components
**External embed support** with security gating via CSP policies

Why it matters for business: If you’re using WebChat for client-facing AI interactions, your agents now present information in a format that looks professional, not like a raw API dump.

3. Teams and Feishu Integration Improvements

Better integration boundaries and configuration for Microsoft Teams and Feishu (Lark), including cleaner message handling and more reliable channel routing.

4. Codex OAuth Stability Fixes

The new Codex provider from 4.10 got immediate attention in 4.11-pre with fixes for OAuth login flows, authentication edge cases, and session persistence. If you adopted Codex in 4.10, 4.11-pre makes it production-ready.

5. Talk Mode and Transcription Fixes

Multiple fixes for audio transcription reliability, Talk Mode stability, and WhatsApp voice message handling — the high-frequency paths that affect daily usability.

Should Your Business Upgrade?

Yes. Here’s the priority breakdown:

Upgrade Immediately If You:

Run coding agents through OpenClaw (Codex integration alone justifies it)
Handle sensitive client data (SSRF hardening is critical)
Use Talk Mode regularly (local MLX voice + stability fixes)
Have ChatGPT history you want to import into OpenClaw’s memory

Upgrade With Caution If You:

Have heavily customized WebChat frontends (rich media changes may affect custom rendering)
Run complex multi-provider setups (test Codex OAuth separately)
Are token-cost sensitive (Active Memory adds per-reply overhead)

How to Upgrade


openclaw update
openclaw gateway restart

After upgrading to 4.10, enable Active Memory with:


openclaw config set memory.active.enabled true
openclaw config set memory.active.mode message

Then update to 4.11-pre and verify Codex OAuth and WebChat rendering before enabling Active Memory’s recent mode.

The Bigger Picture

These releases show where OpenClaw is heading. The platform is evolving from a strong feature stack into what you’d call an operations-grade AI platform: 1. Memory is becoming auditable — not just stored, but traceable, reviewable, and improvable 2. Model providers are getting first-class treatment — Codex, Ollama, local MLX, cloud services all have proper integration paths 3. Security is being hardened proactively — not reactively patching vulnerabilities, but setting strict defaults 4. The UI is becoming a real console — not a chat window with raw text output For businesses running AI assistants on their own infrastructure, this matters. It means OpenClaw is building toward a platform that can handle enterprise workloads — multi-client, multi-provider, with proper security and memory governance. The question isn’t whether to upgrade. It’s whether your team is ready to take advantage of what these updates unlock.

Need help upgrading your OpenClaw instance? Contact All About Web Services — we’ve been deploying and managing OpenClaw installations since the early versions, and we can handle your upgrade from start to finish.

*Last updated: April 13, 2026 | OpenClaw versions 2026.4.10 & 2026.4.11-pre*