Cm

What it does

Cm, or CASS Memory System, provides a procedural memory system for AI coding agents. It transforms scattered session data into persistent, cross-agent knowledge, allowing agents to learn from each other's experiences and avoid repeating past mistakes. The system uses a three-layer architecture (Episodic, Working, and Procedural Memory) to structure and distill information gained across various agents like Claude Code, Codex, Cursor, Aider, PI, Gemini, and ChatGPT.

When to use it

• Before starting non-trivial coding tasks to leverage past experiences and avoid redundant work.
• When encountering recurring bugs or issues that seem to require restarting from scratch.
• To ensure knowledge gained in one agent (e.g., Cursor) is accessible and beneficial to other agents (e.g., Claude Code).
• For cross-agent collaboration, allowing different AI coding tools to share learning and best practices.

Key capabilities

• Cross-Agent Learning: Enables knowledge sharing between different AI coding agents.
• Three-Layer Cognitive Architecture: Organizes memory into Episodic (raw logs), Working (summaries), and Procedural (rules) layers.
• Context Retrieval: Provides relevant rules, anti-patterns, past sessions, and suggested searches based on the current task.
• History Snippet Filtering: Allows filtering of historical data by source ("local" or "remote").
• Confidence Decay System: Rules degrade in confidence over time without revalidation, with a 90-day half-life and a harmful multiplier for mistakes.

Example prompts

• cm context "<your task>" - Retrieves relevant information for the specified task.
• cm context "Debugging authentication errors" - Searches for past solutions to authentication issues.
• historySnippets[].origin.kind is "remote" . origin.host : "workstation" - Filters history snippets to show only remote sessions from a workstation.

Tips & gotchas

• Run the cm context "<your task>" command before starting any significant coding task to benefit from shared knowledge.
• Be aware of the confidence decay system; rules lose effectiveness over time without feedback or revalidation.
• The system relies on session data from various agents, so its usefulness depends on the quality and quantity of available information.