OpenClaw’s 2026.4.29 release adds a set of behavior, memory, provider, and reliability changes across the platform. The update focuses on how messages are routed, how memory is recalled, and how the gateway and channel integrations recover from failures.
On the messaging side, OpenClaw now defaults active-run queueing to steer, with a 500 ms follow-up fallback debounce, according to the release notes. The project also adds a dedicated steering queue docs page, and makes drain behavior more explicit so pending steering messages are handled at the next model boundary. Legacy one-at-a-time steering remains available as queue.
The release also adds visible-reply enforcement controls. Operators can now set a global messages.visibleReplies flag to require visible output to go through message(action=send) for any source chat, while messages.groupChat.visibleReplies remains available as a group or channel-level override, the notes said.
OpenClaw also expands support for opt-in inferred follow-up commitments. The new commitments system uses hidden batched extraction, per-agent and per-channel scoping, heartbeat delivery, CLI management, and a simple commitments.enabled / commitments.maxPerDay configuration. The heartbeat interval now clamps due times so automated check-ins do not echo immediately.
Memory gets a major expansion in the same release. According to the changelog, OpenClaw now treats memory as a people-aware wiki with provenance views, canonical aliases, person cards, relationship graphs, privacy reports, and search modes for person lookup, question routing, source evidence, and raw claims.
Active Memory also gains per-conversation allowedChatIds and deniedChatIds filters. That lets operators limit recall to selected direct, group, or channel conversations while excluding broader sessions.
The release adds partial recall behavior when the hidden memory sub-agent times out. Instead of discarding recovered context, OpenClaw now returns bounded partial recall summaries, including a default temporary-transcript path.
For operators who want diagnostics, OpenClaw now offers a read-only doctor.memory.remHarness RPC to preview bounded REM dreaming output without mutation paths. The gateway also gets an opt-in startup diagnostics timeline that records lifecycle and plugin-load phases behind a config flag.
Model and provider coverage also grows. OpenClaw adds an NVIDIA provider with API-key onboarding, setup docs, static catalog metadata, and literal model-ref picker support. The release also extends Amazon Bedrock support so Claude Opus 4.7 thinking profiles are exposed with xhigh, adaptive, and max modes, while Opus and Sonnet 4.6 stay adaptive by default.
There is also a safety fix for OpenAI Codex-compatible behavior. The notes say OpenClaw now suppresses explicitly configured openai-codex/gpt-5.4-mini inline entries so stale model configs written by doctor --fix cannot bypass manifest capability blocks and trigger repeated assistant-turn failures on ChatGPT-backed Codex accounts.
Reliability work spans the gateway and plugin runtime. The update adds a SQLite-backed plugin state store for restart-safe keyed registries, version-scoped update caches, and startup changes that make slow-host startup easier to diagnose. Gateway watch now runs through a named tmux session by default, and the system emits a startup timeline when diagnostics are enabled.
Channel fixes cover Slack, Telegram, Discord, WhatsApp, Matrix, Feishu, Microsoft Teams, and Signal. The release notes call out Slack Block Kit limits, Telegram proxy and polling resilience, Discord startup and rate-limit handling, and Signal receive-stream behavior that stays open while idle.
Security and operations changes include OpenGrep scanning, sharper GHSA triage policy, safer exec and owner-scope handling, Docker onboarding automation, and an opt-in IPv6 ULA mode for web-fetch behind trusted proxy stacks. The Docker setup also adds OPENCLAW_SKIP_ONBOARDING so automated installs can skip the interactive onboarding step while still applying gateway defaults.
The release includes a long list of fixes as well. Among them are systemd restart-loop protection for occupied ports, better browser control startup behavior, more accurate tool-result overflow checks, safer memory and plugin dependency repair, and multiple channel-specific fixes for empty prompts, delivery retries, and message visibility.
OpenClaw 2026.4.29 also refreshes its dependency stack, updates locale support in the Control UI, and adds documentation entries for Yuanbao channels. The changelog says the release spans messaging, memory, providers, gateways, channels, security, and runtime repair across the full stack.
