Agent Memory Guard Audrey 1.0 Released

Audrey is a local-first memory/control layer that controls AI agent behavior and manages memory. It adopts a 'memory-before-action' approach where the model proposes an action and the host decides on it, functioning as substantive control infrastructure at the tool boundary, beyond simple system prompt advice.

Audrey 1.0 Key Updates:

• Pre-action Judgment: Provides allow, warn, and block decision capabilities before execution
• Memory Security: Supports redacted memory for tool-trace records
• Benchmark: Includes the GuardBench benchmark and artifact bundle
• Extensibility: Enhanced MCP (Model Context Protocol) support and Node/Python clients provided

This tool focuses on practical problem-solving that increases the reliability and safety of agent operations, such as preventing repetition of destructive commands, warning based on previous fixes, detecting outdated schema assumptions, and identifying retry loops using the same strategy.