OpenClaw v2026.9.4 Released, Introducing Conversation-Based Skill Learning

OpenClaw v2026.9.4 was released with the participation of 1,558 PRs and 293 contributors. This update improves plugin and skill discovery and introduces a feature to learn reusable skills from past conversations.

Conversation-Based Skill Learning

Through the 'Learn from past conversations' feature in 'Skill Workshop', users can generate skills based on past conversations. The process is conducted via a chat interface, allowing users to steer or stop the learning direction. Auto mode immediately applies improvements, while Propose mode suggests them in a pending approval state. Model usage costs are billed normally regardless of whether auto-learning is enabled.

GPT Image 2.5 and Cloud Session Management

GPT Image 2.5 Flare and Sunburst models are now supported, enabling custom sizes, high-quality settings, transparent background output, and editing via reference images. Separate OpenAI API keys or fal credentials are required. In terms of cloud session management, administrators can save prepared projects and snapshots for reuse, and manage standby instances for the next session via the Ready workers setting. The default setting is 1 standby instance per project, with a total shared limit of 4.

Terminal Interaction and Other Improvements

Claw can now present choices or request text input from users in the terminal. Masking is applied when entering sensitive information so it does not remain in the conversation history. Other bug fixes include Discord voice memo processing, Feishu document tool restoration, improved memory search accuracy, and prevention of update stalls.