OpenClaw v2026.5.22 Release: 4,100x Performance Boost, Meeting Notes Plugin, and Hundreds of Fixes

The core question this article answers: What truly matters in OpenClaw’s latest release? Why is this version a milestone for Gateway performance and real-world usability?

On May 22, 2026, the OpenClaw team released v2026.5.22. This is far more than a routine monthly update—from a radical Gateway performance leap to the first external “Meeting Notes” plugin, plus over two hundred fixes and optimizations, this release redefines the daily experience for development and automation. If you use OpenClaw to build multi‑channel AI agents or struggle with slow model listing, stuck sessions, or confusing sub‑agent configuration, this update directly addresses your pain points.

📌 One‑Minute Summary

• ❀
  
  Massive performance gains: /models calls drop from 20 seconds to 5 milliseconds (~4,100× faster). Extensive caching and lazy loading make Gateway startup and runtime much lighter.
• ❀
  
  Brand‑new Meeting Notes plugin: First source‑only external plugin supporting live Discord voice capture, manual transcript imports, and read‑only CLI access (openclaw meeting-notes).
• ❀
  
  Hundreds of fixes: Covers Telegram, Discord, Codex, WhatsApp, Windows installer, session locks, OAuth refresh, proxy bypass, and many other components.
• ❀
  
  Security & packaging improvements: Smaller npm package (excludes docs images), shrinkwrap locking for dependency graphs, and plugin SDK enhancements (poll sender, row‑level session helpers).
• ❀
  
  Better developer experience: New openclaw-landable-bug-sweep skill for batch PR generation, improved QA‑Lab observability (Prometheus/OTel), and Windows “no‑environment” auto‑setup.

1. Gateway Performance Revolution: From 20 Seconds to 5 Milliseconds

1.1 Core problem: Why does listing models take half a minute?

In multi‑provider setups (e.g., switching between OpenAI, Anthropic, xAI), many users complained that /models or model auto‑completion was extremely slow. The root cause: every request triggered file stats, manifest reloads, and even external CLI probes for every provider plugin. With several plugins running locally, overhead accumulated to 10–20 seconds.

1.2 The solution: Pre‑warming + immutable snapshots + lazy loading

OpenClaw v2026.5.22 introduces a combination of techniques:

Real‑world example: On a Gateway configured with six model providers (including local Ollama, cloud Codex, Groq, Copilot), openclaw models list dropped from ~22 seconds to ~5 milliseconds. Cold start time decreased by about 40% because unused handler trees are no longer loaded.

“

Reflection: This 4,100× speedup teaches a valuable lesson – laziness must be context‑aware. For metadata that is frequently read and rarely changes, pre‑warming at startup is far smarter than runtime probing. Many teams default to “lazy‑load everything”, but data like the model catalog, which can be queried at any time, greatly benefits from eager loading.

2. Meeting Notes Plugin: Let AI Capture Your Discord Voice Discussions

2.1 Core problem: How can I automatically generate meeting minutes from multi‑person voice chats?

Team collaboration, remote meetings, client calls – we often need to manually transcribe and summarise discussions. OpenClaw v2026.5.22 introduces the Meeting Notes plugin to close this loop.

2.2 Features explained

• ❀
  
  Source‑only external plugin: Plugin code lives outside the core npm package and adheres to a source‑provider contract in the SDK. This makes it easy to add new live sources (first is Discord voice).
• ❀
  
  Automatic capture: Once configured, OpenClaw automatically transcribes and summarises whenever someone speaks in a Discord voice channel.
• ❀
  
  Manual import: Upload existing transcript files from other meeting tools.
• ❀
  
  Read‑only CLI access: openclaw meeting-notes lets you safely browse history without accidental modifications.
• ❀
  
  Future extensibility: The architecture allows adding Zoom, Google Meet, and other sources later.

2.3 Use case scenario

Scenario: Weekly remote team sync

• ❀
  
  Team uses Discord voice for weekly meetings.
• ❀
  
  Enable meeting-notes plugin in OpenClaw config and bind a Discord account.
• ❀
  
  When the meeting starts, the agent automatically joins the voice channel, transcribes speech in real‑time, and produces structured notes (decisions, action items, risks).
• ❀
  
  After the meeting, team members run openclaw meeting-notes list --date=2026-05-22 to retrieve a Markdown summary.

This plugin marks the first time “conversation‑as‑document” is brought into OpenClaw, fully open source, so you can customise transcription engines or output formats.

“

Reflection: I particularly admire the “source‑only” design – instead of cramming the feature into the core, the team provides a clean SDK contract and encourages the community to build adapters for different platforms (Teams, Zoom, Feishu). This “stable kernel + flexible periphery” pattern is a hallmark of successful open source projects.

3. Other Notable Features & Improvements

3.1 Sub‑agent context pruning: safer and more efficient delegation

Core problem: When I invoke a sub‑agent, it always inherits my full persona, identity, memory files, causing token explosion and privacy leaks.

Solution: Sub‑agents now inherit only AGENTS.md and TOOLS.md by default, excluding persona, identity, user, memory, heartbeat, and various setup files. This drastically reduces startup tokens and prevents sub‑agents from reading sensitive memory.

Use case: Suppose your main agent holds a large amount of personal diary memory. You need a sub‑agent to search public web information. In older versions, the sub‑agent would load all private memories, risking accidental leakage. Now it receives only the tools and basic behaviour files – clean and safe.

3.2 CLI onboarding: classic wizard runs when config is missing

Core problem: New users install OpenClaw, type openclaw, and have no idea what to do next.

Solution: When no config.json exists, plain openclaw now triggers classic onboarding (step‑by‑step provider and channel configuration). Already‑configured installations continue with Crestodian (advanced config management). In short: ask first, don’t crash.

3.3 Discord component lifecycle configurable

Core problem: With interactive components (buttons, dropdowns) in Discord, callback registrations stayed in memory forever, causing memory leaks over long runs.

Solution: New agentComponents.ttlMs setting lets you define a maximum lifetime for callbacks per account (max 24 hours). Expired callbacks are automatically cleaned up.

3.4 xAI/Grok deep integration

• ❀
  
  Reuse xAI OAuth profiles for Grok’s web_search tool.
• ❀
  
  Add Grok model aliases (e.g., grok-3-latest).
• ❀
  
  Media providers can declare default operation timeouts, preventing long‑running image/video generations from blocking.

3.5 Plugin SDK enhancement: row‑level session helpers

Core problem: Plugins that needed to read or write session data previously used loadSessionStore, which loaded the entire store – inefficient and prone to conflict.

Solution: Introduce row‑level session workflow helpers, allowing plugins to read or patch specific session fields precisely. The old API is deprecated. This is a big step toward fine‑grained permissions and concurrency safety.

3.6 Packaging & security: shrinkwrap + trimmed assets

• ❀
  
  The npm package now includes a generated shrinkwrap, and any lockfile/shrinkwrap changes require review. This ensures every user install of OpenClaw and its official plugins uses the exact same dependency graph, preventing malicious package substitution.
• ❀
  
  Documentation images and assets are excluded from the npm tarball, reducing package size by about 30% for faster downloads.

4. Important Fixes: Potholes You May Have Hit – Now Filled

This release fixes over 150 issues. Here are 12 typical fixes in a table:

Besides the table, several fixes deserve extra detail:

4.1 Session locks and orphaned processes

• ❀
  
  ACP child session leaks: Previously, when a parent session was reset, claude-agent-acp child processes spawned via sessions_spawn never closed, gradually exhausting memory. Now deleting a parent also closes child ACP sessions. (#85190)
• ❀
  
  Session write‑lock watchdog: If auto‑compaction stalled, it held the lock for the entire run window. Now compaction timeout is independent, preventing long lock hold. (#84949)

4.2 Installation & update robustness

• ❀
  
  Windows bare‑metal setup: If raw Windows has no Node, winget, Chocolatey, or Scoop, the installer now downloads a user‑local portable Node.js, extracts it using built‑in tar or .NET fallback, and continues installation. This dramatically lowers the entry barrier for Windows.
• ❀
  
  Git update rollback: If git pull followed by pnpm install/build fails, the system rolls back to the previous checkout, avoiding a broken installation.
• ❀
  
  macOS LaunchAgent race conditions: Fixed several cases where concurrent bootstrap caused the service to be unregistered or falsely report failure; now waits for launchd to fully complete.

4.3 Multi‑channel message handling

• ❀
  
  Telegram code blocks: Preserve language-* classes when sending to Telegram, enabling proper syntax highlighting.
• ❀
  
  Slack reasoning content hidden: Suppress all reasoning payloads to prevent internal model thoughts from leaking to users.
• ❀
  
  Discord component cleanup: Old cleanup callbacks no longer mistakenly remove newer registrations for the same account when reconnecting.
• ❀
  
  Matrix two‑person room detection: Now favours explicitly configured rooms over m.direct, preventing mention gating from being bypassed.

5. Author’s Reflection & Unique Insights

“

As a long‑time OpenClaw user for internal automation, the most striking part of this release is not a single feature but the methodology of performance optimisation and the pattern of open‑source collaboration.

On performance: Many people think optimisation means “add caching, change algorithms”. The OpenClaw team did one thing right first: measure. They found that 95% of the time in /models was wasted on repeated plugin discovery and filesystem probes – results that are almost immutable at runtime. By applying “pre‑warming + immutable snapshots”, they achieved a 4,000× improvement without any complex parallelism or algorithm rewrites. This reminds us: performance bottlenecks are rarely where you guess; profile first, then eliminate duplicate work.

On the plugin system: The Meeting Notes plugin is released as “source‑only” outside the core package, with a clear SDK source‑provider contract. This means anyone can add an adapter for Zoom, Feishu, or even local audio files without modifying a single line of OpenClaw core. This “minimal core + maximal periphery” architecture has proven successful in projects like Kubernetes and VS Code. OpenClaw is evolving from a mere chatbot framework into a pluggable conversation‑driven automation operating system.

On community collaboration: This release thanks over 50 contributors. Many documentation fixes and platform‑specific compatibility patches (especially Windows and macOS edge cases) came from community users who submitted PRs. For example, the Windows portable Node.js support was driven by a developer’s pain point with a “no winget” environment. This again shows that an active community is the lifeblood of open source, and maintainers who listen to pain points and turn them into code are what make projects thrive.

6. Practical Summary & Action Checklist

If you plan to upgrade to v2026.5.22, follow these steps for the best experience:

1. Back up configuration and state

   cp -r ~/.openclaw ~/.openclaw.backup
   

2. Upgrade the package

   npm update -g openclaw
   # or for git install
   git pull && pnpm install && pnpm build
   

3. Run doctor to automatically migrate configs

   openclaw doctor --fix
   

   This updates old Groq, GitHub Copilot, Claude, etc., model references to current provider catalogs.

4. Test model list performance

   time openclaw models list
   

   Expect a drop from seconds to milliseconds.

5. Enable Meeting Notes plugin (optional)

   • ❀
     
     Install plugin: openclaw plugins install openclaw-meeting-notes
   • ❀
     
     Configure Discord voice source (see plugin documentation)
   • ❀
     
     Run openclaw meeting-notes capture --source discord --channel "general"

6. Verify sub‑agent behaviour
   If you use custom sub‑agents, check that they only inherit AGENTS.md and TOOLS.md and adjust as needed.

7. Windows users specific steps

   • ❀
     
     If a previous installation fails, try a clean run (the installer will automatically handle portable Node.js).
   • ❀
     
     Use cmd.exe or PowerShell, not WSL2 UNC paths, to start the CLI.

One‑Page Summary

7. Frequently Asked Questions (FAQ)

Q1: After upgrade, my custom sub‑agent config files (like persona.md) are no longer loaded. What should I do?
A: The new version excludes persona, identity, memory, etc., by default. If you need the sub‑agent to inherit these, explicitly add inheritFiles: ["persona.md"] when calling sessions_spawn.

Q2: Does the Meeting Notes plugin support Zoom or Teams?
A: The current release only includes Discord voice as a live source. However, thanks to the source‑provider SDK, the community can quickly add other platforms. Watch the openclaw-meeting-notes plugin for future updates.

Q3: Does the performance optimisation make Gateway startup slower because of pre‑warming?
A: Pre‑warming happens asynchronously during startup and does not block health checks or readiness signals. In fact, overall startup is faster because lazy loading avoids unused components.

Q4: Will shrinkwrap lock dependencies of third‑party plugins and cause installation failures?
A: Shrinkwrap only applies to the core OpenClaw package and its official npm plugins. Third‑party plugins can still declare their own dependencies; pnpm resolves them without conflicts.

Q5: I still get pnpm: command not found on Windows. What should I do?
A: The new installer places a portable Node in the user directory and adds it to PATH. If an existing installation has issues, delete %LOCALAPPDATA%\openclaw and rerun the installer.

Q6: Can the 24‑hour limit for Discord component callbacks be adjusted?
A: Yes. Set agentComponents.ttlMs per account under channels.discord configuration. The maximum is 86400000 (24 hours); values beyond that are capped.

Q7: What’s improved with Codex OAuth refresh in this release?
A: Previously, OAuth credentials without a refresh token would hang for 120 seconds before failing. Now they fail immediately with a hint to run openclaw doctor to re‑authorise. Also, deactivated_workspace is classified as a permanent failure, triggering fallback models.

Q8: Can I run this version in Docker and use macOS VM auto‑login?
A: Yes, the documentation includes updated guidance for macOS VM auto‑login. The Docker image now pre‑creates workspace and auth‑profile mount points with the correct node ownership, avoiding permission issues on first run.

OpenClaw v2026.5.22 is not just a version bump – it’s a deep polish of performance, security, and usability. Whether you’re an individual developer, an enterprise automation team, or an open‑source contributor, you’ll enjoy a smoother, more reliable AI agent experience. Upgrade today and enjoy millisecond model switching and automated meeting notes!

“

Image credits: Unsplash (https://unsplash.com/photos/computer-code-on-screen-2y2KkXqjVrY) and Pexels (https://www.pexels.com/photo/white-and-black-printer-paper-16417605/) for illustrative purposes only.