OpenClaw v2026.4.29 Released: Memory upgraded to "Character Wiki," agents learn to proactively track commitments

According to Beating Monitoring, the open-source personal AI assistant OpenClaw (GitHub stars: 367K) released version 2026.4.29, marking its second update in two days.
The biggest changes for users are in the memory system and messaging system.

Memory has evolved from simple retrieval-based recall to a character perception wiki: the agent can automatically build character profiles, track interpersonal relationship graphs, with each memory source traceable and labeled with evidence type. Active Memory now supports filtering by conversation ID, allowing operators to enable memory recall only for specified conversations; sub-agents no longer discard retrieved content after timeout, but instead return partial recall summaries.

The messaging system adds a “Promise” feature. Agents implicitly extract follow-up promises during conversations (such as “Remind you tomorrow” or “Check progress next week”), which are delivered periodically via a heartbeat mechanism, with configurable limits per agent and channel. The queuing strategy for user messages has also changed: it now defaults from sequential queuing to steer mode, where all pending messages are read at once within the model inference boundary to reduce interruptions.

Other changes include: NVIDIA joining as a built-in model provider, callable with an API key; plugin updates adding SQLite persistent state storage that retains data across restarts; security layer integration of OpenGrep rule scanning with automatic upload of SARIF reports to GitHub Code Scanning; plugin loading time on Windows reduced from about 39 seconds to approximately 2 seconds; and over ten Block Kit overflow issues in Slack channels have been fixed.