OpenClaw adds a new Active Memory plugin, automatically retrieving memories during conversations without manual triggers

ME News message: On April 13 (UTC+8), according to monitoring by 1M AI News, the open-source AI Agent tool OpenClaw released version v2026.4.12. The key highlight is the addition of the Active Memory plugin.

Active Memory is an optional plugin that automatically runs a dedicated memory sub-agent before each reply. It proactively retrieves user preferences relevant to the current conversation, historical context, and past details—without requiring the user to manually input “remember this” or “search memory.”

The plugin supports three context modes (single message, recent messages, full conversation), and allows you to view the memory retrieval process in real time via the /verbose command.

Other major new features:

1. macOS Talk Mode adds an experimental local MLX speech engine. Speech synthesis can run locally without calling cloud APIs, and will automatically fall back to the system speech when unavailable.
2. A built-in Codex model provider enables \codex/gpt-* models to be used directly via native Codex authentication and thread management.
3. A built-in LM Studio provider supports locally hosted and self-managed OpenAI-compatible models, including automatic model discovery and embedded vector search.
4. Added the openclaw exec-policy command, which allows you to manage execution approval policies locally.
5. Plugin loading optimization: at startup, it loads only the necessary components declared in the manifest, avoiding loading unrelated plugins during runtime.

Security fixes addressed three vulnerabilities: preventing authorization-check bypass via empty approval lists; removing busybox/toybox from the secure binary list to prevent them from being used for interpreter escape; and enhancing shell wrapper detection to block environment variable injection attacks.

(Source: BlockBeats)