OpenClaw 4.26: New AI Voice + Local Updates

OpenClaw 4.26 is presented as a significant update primarily focused on fixing months of frustration around local AI model support. The Ollama integration received a comprehensive overhaul addressing nearly every known pain point: model name prefixes are now stripped cleanly before requests, network discovery only runs when explicitly opted in, custom remote Ollama setups (including cloud-hosted ones) connect properly, timeouts follow user configuration, thinking controls map correctly to Ollama's native format, tools register based on actual model capabilities, context windows respect model settings instead of defaulting to maximum VRAM usage, and memory embeddings now use the correct `/api/embed` endpoint with batched input. The net result is that local models use less memory, respond faster, and produce fewer mysterious errors. These improvements extend beyond Ollama to all local OpenAI-compatible providers including LM Studio, vLLM, and SG Lang, with loopback connections trusted by default and a new diagnostic that clearly reports when a local model runs out of RAM.

A new migration tool addresses one of the biggest barriers to tool adoption. Running `openclaw migrate` handles the full transfer of configuration, memory settings, model providers, MCP server connections, skills, commands, and credentials from both Claude Code and Hermes agent setups. It includes a dry-run mode, creates a backup before making changes, and is designed to let users keep their existing workflows while gaining OpenClaw's additional capabilities like multi-channel messaging, voice agents, and browser automation.

The voice capabilities in 4.26 expand meaningfully with Google Live voice sessions now working directly in the browser through talk mode, powered by Gemini live two-way audio with tool access mid-conversation. The agent consult feature from the previous release is also supported here, allowing a voice agent to pause, query the full OpenClaw agent, and return with a substantive answer. A new backend relay for voice plugins enables server-side audio processing, useful for business phone line scenarios where the browser shouldn't have direct access to the voice provider.

Other notable additions include one-command end-to-end encryption setup for Matrix messaging, full QQ group chat support for agents serving Chinese markets, Tencent's Yuanbao platform added to the official channel catalog, and Cerebras added as a new bundle provider for fast AI chip-based inference. Memory search accuracy improved for specific embedding models through proper query prefix formatting, and a new asymmetric embeddings configuration option fixes setups that previously produced poor results.

The compaction system gained a file-size-based safety trigger to complement the existing token-count trigger, and a fix ensures compaction summaries are rebuilt from actual conversation content rather than stacking on previous summaries. Log redaction now extends to session transcripts. Long-standing session reset bugs caused by background tasks, heartbeat checks, and cron jobs keeping sessions artificially alive have been resolved, with background activity now tracked separately from real user activity. Cron job reliability improved with interrupted jobs properly recorded as failed and one-shot jobs protected from double execution. The install and update process was hardened with atomic NPM updates, a Docker home directory permissions fix, a macOS launch agent split-brain detection fix, and browser automation improvements that prevent Chrome from spawning in infinite crash loops and clean up stale tabs from previous sessions.