🦞 OpenClaw v2026.4.12

"Quality & Reliability Release"

Released: April 13, 2026
Focus: Plugin Loading, Memory, Dreaming, Local Models
Generated: April 14, 2026

📊 Release Overview

Major focus: Stability, security hardening, and developer experience

✨ Headline Features (1/3)

🧠 Active Memory Plugin

• Configurable message/recent/full context modes
• Live /verbose inspection
• Advanced prompt/thinking overrides
• Opt-in transcript persistence

✨ Headline Features (2/3)

🤖 Codex Provider Integration

💻 LM Studio Provider

• Onboarding and runtime model discovery
• Stream preload support
• Memory-search embeddings
• Local/self-hosted OpenAI-compatible models

✨ Headline Features (3/3)

🍎 MLX Speech Provider (macOS)

🔧 Exec Policy Management

• openclaw exec-policy command
• Show, preset, and set subcommands
• Sync tools.exec.* config with local approvals
• Node-host rejection and rollback safety

🔌 Plugin System Improvements

• Narrowed activation scope — CLI, provider, and channel activation limited to manifest-declared needs
• Trust boundaries — Explicit scope preservation
• Centralized manifest policy — Startup, command discovery, and runtime activation avoid loading unrelated plugins
• Commands RPC — commands.list RPC for remote gateway clients to discover runtime-native, text, skill, and plugin commands

🧠 Memory & Dreaming Enhancements

Active Memory

• Default QMD recall to search with better telemetry
• Keep recall runs on resolved channel
• Improved lexical fallback ranking

Dreaming

• Consume managed heartbeat events exactly once
• Stage light-sleep confidence from all short-term signals
• Wake scheduled jobs immediately
• Stop dreaming from re-ingesting own narrative transcripts

🔒 Security Hardening

• CRITICAL: Remove busybox/toybox from interpreter-like safe bins
• CRITICAL: Prevent empty approver list from granting explicit approval authorization
• CRITICAL: Broaden shell-wrapper detection and block env-argv assignment injection
• Blank shipped example gateway credentials in .env.example
• Fail startup when placeholder tokens are still configured

📡 Gateway & Channel Improvements

Gateway Core

• Split startup and runtime seams for better lifecycle sequencing
• Defer scheduled services until sidecars finish
• Gate chat history during sidecar resume
• Load live gateway slash-command catalog into composer

Matrix

• MSC4357 live markers for draft preview sends
• Typewriter animation support
• Keep room mention gating strict with @displayName labels

📱 Platform-Specific Improvements

WhatsApp

• Centralize per-account connection ownership
• Fall back to first mediaUrls entry when mediaUrl is empty

iMessage

• Retry transient watch.subscribe startup failures
• Sanitize startup error logging

Telegram

• Route approval button callbacks onto separate sequentializer lane
• Live openclaw qa telegram lane for bot-to-bot checks

🧪 QA & Testing Infrastructure

• Multipass runner — --runner multipass for repo-backed QA scenarios in disposable Linux VMs
• Telegram QA — Live private-group bot-to-bot checks with artifact handling
• Convex-backed credential leasing — Pooled Telegram credential leasing with openclaw qa credentials admin commands
• Packaging fixes — Stop packaged QA helpers from crashing when optional scenario config is unavailable

🛠️ Developer Experience

• Control UI improvements — Simplified Scene and Diary surfaces, stable waiting-entry ordering
• Doctor enhancements — Warn when on-disk agent directories exist but config entries are missing
• CLI updates — Respawn tracked plugin refresh after package self-updates
• Docs i18n — Chunk raw doc translation, reject truncated outputs, recover from terminated Pi sessions
• Provider docs — Expanded bundled provider docs with richer capability, env-var, and setup guidance

📚 Memory Wiki Improvements

• Unicode support — Preserve Unicode letters, digits, and combining marks in wiki slugs
• Contradiction clustering — Cap Unicode filename segments to safe byte lengths
• Short-term recall — Allow nested daily notes under memory/**/YYYY-MM-DD.md
• Dreaming exclusion — Exclude generated dream reports under memory/dreaming/**
• QMD + bridge-mode — Recommended hybrid recipe with zero-artifact troubleshooting

🐛 Major Bug Fixes (1/2)

• Gateway startup — Defer scheduled services until sidecars finish, preventing race conditions
• Agent queueing — Carry orphaned active-turn user text into next prompt before repairing transcript ordering
• Gateway keepalive — Stop marking WebSocket tick broadcasts as droppable to prevent self-disconnect
• Telegram approval — Route approval button callbacks onto separate lane to prevent deadlocks
• Discord gateway — Clear stale heartbeat timers before reconnecting to prevent zombie callbacks

🐛 Major Bug Fixes (2/2)

• Dreaming narrative — Harden transient narrative cleanup with retry logic and lock-aware session-store path
• Memory-core dreaming — Keep bundled memory-core loaded alongside external memory slot owner only when dreaming is enabled
• Audio transcription — Fix multipart FormData fields being silently dropped, restoring OpenAI audio transcription
• Anthropic replay — Preserve immutable signed-thinking replay safety, drop conflicting tool IDs before validation
• WhatsApp connection — Centralize per-account connection ownership to fix reconnects and login recovery

🔄 Migration Guide

Upgrade Steps

# Update OpenClaw
cd ~/.openclaw
git pull
git checkout v2026.4.12

# Reinstall dependencies
npm install

# Run doctor to check configuration
openclaw doctor

# Restart gateway (if running)
openclaw gateway restart

# Verify plugin loading
openclaw plugins list

⚠️ Breaking Changes & Considerations

• Plugin loading scope — Narrowed to manifest-declared needs; may affect custom plugins
• Gateway credentials — Startup fails with placeholder tokens; update .env before launch
• Memory-core dreaming — Bundled memory-core loading behavior changed when external memory slot owner is present
• Exec policy — New openclaw exec-policy command for managing tool execution approvals

🐛 Known Issues & Workarounds

• macOS + Node 25 — Recursive memory/**/*.md glob watching may fail; memory-core now watches memory directory directly
• Windows packaged installs — May need to reinstall bundled runtime packages if platform native optional child is missing
• Discord preview-streaming — openclaw doctor --fix may rewrite legacy config; manual edit may be needed for downgrades
• QMD scope — Default QMD scope now allows channel sessions but still denies groups

⚡ Performance Impact

Improvements

• Plugin loading — Narrowed scope reduces startup time and memory footprint
• Memory recall — Better lexical fallback ranking improves recall accuracy
• Gateway startup — Deferred scheduled services prevent race conditions

Considerations

• Active Memory — Adds sub-agent run before main reply; configurable context modes allow tuning
• Dreaming — Improved confidence scoring may affect promotion timing

👥 Community Response

• Security hardening welcomed — Three critical fixes address privilege escalation concerns
• Active Memory praised — Automatic context recall reduces manual "remember this" commands
• Local model support significant — LM Studio and MLX providers enable offline/private deployments
• Plugin loading improvements — Narrowed scope reduces startup complexity
• Stability focus appreciated — Bug fixes address long-standing gateway and dreaming issues

⚔️ Competitive Positioning

vs Hermes Agent

• Memory systems — OpenClaw's Active Memory vs Hermes' Honcho integration
• Local models — OpenClaw adds LM Studio, Hermes has broader provider support
• Platform focus — OpenClaw multi-channel, Hermes 16 platforms

Unique Strengths

• Dreaming system — Unique long-term memory consolidation
• QA infrastructure — Multipass runner and Telegram bot-to-bot testing
• Plugin ecosystem — Mature plugin system with trust boundaries

🗺️ Roadmap Implications

• Quality focus continues — This release signals ongoing stability prioritization
• Local-first direction — LM Studio and MLX providers indicate privacy/offline focus
• Memory evolution — Active Memory plugin suggests continued memory system innovation
• Plugin maturity — Narrowed loading scope indicates plugin ecosystem maturation
• Security hardening — Three critical fixes show proactive security posture
• Developer experience — Control UI and docs improvements signal UX focus

👨‍💻 Developer Impact

Positive

• Better plugin loading reliability
• Improved memory and dreaming stability
• Local model options (LM Studio, MLX)
• Enhanced QA infrastructure
• Security hardening

Considerations

• Plugin manifest updates may be needed
• Gateway credential validation now enforced
• Active Memory adds sub-agent overhead

🎯 Key Takeaways

1. Quality release — Focus on stability, reliability, and bug fixes
2. Security hardening — Three critical security fixes
3. Memory innovation — Active Memory plugin for automatic context recall
4. Local-first — LM Studio and MLX providers enable offline/private deployments
5. Plugin maturity — Narrowed loading scope with trust boundaries
6. Developer experience — Improved Control UI, docs, and QA infrastructure

📚 Resources

• Release Notes: GitHub v2026.4.12
• Documentation: docs.openclaw.ai
• Discord Community: discord.gg/openclaw
• GitHub Repository: github.com/openclaw/openclaw