OpenClaw introduces the "Dream" memory system, which shares multiple similarities with Claude Code's leaked source code

According to 1M AI News monitoring, open-source AI assistant platform OpenClaw released v2026.4.5, featuring about 40 new features and over a hundred fixes, delivered by more than 60 contributors. This is the first major version update since Anthropic banned OpenClaw subscription access.

The most eye-catching new feature is an experimental “dreaming” memory system that splits backend memory integration into three cooperating stages: light sleep filters short-term memories, deep sleep handles durable persistence writes, and the REM stage discovers relationships and distills lasting insights. Each stage has its own independent scheduling and can run automatically when no user is interacting, along with a Dream Diary interface and debugging tools. In the Claude Code source code that was accidentally leaked on March 31, there happens to be a mechanism named autoDream as well—also automatically integrating, deduplicating, and eliminating memory conflicts during idle time, using a similar layered architecture: the index is always loaded, topics are read on demand, and history is only searched without loading.

Another similarity is prompt cache optimization. In this version, OpenClaw improves cache hit rates by deterministically ordering MCP tool definitions, normalizing system prompt fingerprints, stripping inline tool lists, and other measures. The Claude Code leaked source code shows that its internal design uses stable/dynamic prompt boundaries to maximize cache reuse; the optimization directions are nearly identical. Claude Code engineer Boris Cherny himself is also one of the contributors to OpenClaw’s cache improvements in this release.

However, the development records of relevant OpenClaw PRs indicate that these features were built well before the leak—finishing such a scale of feature migration in under a week is not realistic from an engineering perspective. A more likely explanation is that the two projects arrived at similar approaches when solving the same problem, and the leak gave the outside world the first chance to directly compare the internal designs of both.

Other major updates include:

1. Built-in video generation tools, supporting xAI, Alibaba Cloud Wan, and Runway; music generation tools support Google Lyria and MiniMax
2. Added model providers such as Tongyi Qianwen, Fireworks AI, and StepFun; Amazon Bedrock added Mantle for auto-discovery
3. The control panel adds 12 languages including Simplified Chinese, Traditional Chinese, Japanese, and Korean
4. Multiple security hardenings for Claude CLI integration, clearing inherited environment variables and configuration overrides, enforcing restricted permission mode, and preventing third parties from hijacking the Claude authentication context through OpenClaw