OpenClaw 2026.4.7 adds a headless inference hub, memory-wiki, and webhook TaskFlows

• OpenClaw's release post and the release notes frame v2026.4.7 around five headline changes: openclaw infer, media editing upgrades, session branch and restore, webhook-driven TaskFlows, and the new memory-wiki layer.
• According to the infer announcement and the infer docs, openclaw infer is a single headless CLI surface for model runs, image work, audio, TTS, video, web search, and embeddings.
• The memory-wiki thread matches the memory-wiki docs: OpenClaw now ships a structured knowledge vault with claim and evidence fields, contradiction tracking, freshness signals, compiled digests, and optional Obsidian-friendly output.
• The TaskFlow thread and the Task Flow docs position webhook ingress as a way for outside systems to create and drive durable multi-step flows through authenticated per-route endpoints.
• Composio's OpenClaw launch and its security post extend the release story beyond OpenClaw itself: a separate CLI now handles app auth flows that Composio says are often left in plaintext on local agent machines.

You can skim the full release notes, jump straight into the infer command tree, and inspect how memory-wiki splits durable knowledge from ordinary active memory. The other useful rabbit hole is Task Flow, which already had managed and mirrored modes before this release added webhook ingress. Composio also picked the same week to ship a CLI for OpenClaw-style agent auth, with a blog post built around the blunt point that too many agents still sit next to plaintext tokens.

The new infer surface is the cleanest part of this release. Instead of exposing raw provider APIs or tool IDs, the docs describe openclaw infer as a canonical headless interface organized by capability family.

That command tree currently covers:

• model run
• image generate
• audio transcribe
• tts convert
• video generate
• web search
• embedding create

The release notes add two details the launch tweet only hints at: infer is meant for provider-backed workflows across model, media, web, and embeddings, and it is packaged as a stable CLI surface for scripting and skill-building rather than a thin wrapper over gateway internals.

OpenClaw's best new idea is memory-wiki. The plugin docs describe it as a bundled layer beside active memory, not a replacement for it: active memory still handles recall, promotion, indexing, and dreaming, while the wiki compiles durable knowledge into deterministic pages and machine-readable digests.

The mechanics are more opinionated than a folder of notes:

• structured claims and evidence metadata
• page-level provenance and confidence
• contradiction and freshness tracking
• open questions and dashboards
• compiled digests for runtime consumers
• vault modes including isolated, bridge, and local variants

The Obsidian-friendly export option from the launch thread makes the whole thing feel half agent memory system, half maintained internal wiki.

Task Flow already existed as OpenClaw's durable orchestration layer. The docs define it as the substrate above background tasks, with revision tracking, persistent state, and two sync modes: managed flows that own execution, and mirrored flows that watch external work.

What v2026.4.7 adds is ingress. OpenClaw's thread says authenticated webhooks can now create and drive bound TaskFlows over per-route shared-secret endpoints, which turns the flow system into something outside schedulers, services, or SaaS triggers can poke directly.

The release notes also bundle a quieter set of operational upgrades. Media generation now falls back across authenticated image, music, and video providers by default, while preserving intent and remapping size, aspect, resolution, or duration hints to the nearest supported option. The same notes call out mode-aware video-to-video support.

Session handling got branch and restore controls in the headline tweet, which is exactly the kind of small feature that becomes a big deal once OpenClaw is running as a long-lived local harness instead of a throwaway CLI session.

That local-first audience is visible in both directions. bridgemindai's post shows OpenClaw running on an NVIDIA DGX Spark after rate-limit frustration with Claude Code, while one user report says they downgraded to v2026.4.2 because the update_plan tool was causing problems.

Composio's launch landed one day earlier, but it fits this story because it answers a different bottleneck: credentials. The CLI docs frame the tool as a workflow layer around login, search, execute, link, run, and proxy commands, with composio login opening a browser auth flow and composio link connecting external accounts like GitHub, Gmail, or Slack.

The sharper claim is in Composio's blog post. It argues that many agent setups still leave API keys and OAuth tokens in plaintext config files on local machines, then uses March's Mercor, Axios, and other supply-chain incidents as the backdrop for why OpenClaw-style always-on agents need a separate auth layer at all.