mmag

What This Skill Does

MMAG (Mixed Memory-Augmented Generation) is a cognitive architecture for OpenClaw AI agents that transitions them from stateless responders to persistent, long-term companions. By utilizing five distinct memory layers—Conversational, Long-Term User, Episodic, Sensory, and Working—the MMAG skill allows an agent to synthesize past interactions with current situational context. It functions by organizing data into specific layers, ensuring that your agent doesn't just remember what was said, but understands the significance of preferences, time-bound events, and environmental factors. When initialized, the skill provides a unified context window that prioritizes user traits and episodic events, ensuring the agent remains coherent, personalized, and efficient across multiple sessions.

Installation

To integrate MMAG, ensure you have the OpenClaw CLI installed, then execute the following command in your terminal:

1. Install via ClawHub: clawhub install openclaw/skills/skills/j0ker98/mmag
2. Initialize the environment: ./~/.openclaw/skills/mmag/init.sh
3. Integration: Every session must begin by executing ./~/.openclaw/skills/mmag/context.sh and passing its output into your agent's system prompt to activate the memory layers.

Use Cases

MMAG is ideal for users building highly personalized AI assistants. Use it for:

• Personal Health Assistants: Tracking medication schedules (episodic) and wellness preferences (long-term).
• Professional Project Managers: Maintaining current task status (working) while remembering project deadlines (episodic) and stakeholders' specific communication preferences (long-term).
• Travel Companions: Adjusting tone based on the user's current location and local weather (sensory) while tracking flight itineraries and past destination experiences (episodic).

Example Prompts

1. "I'm starting a new project focused on Python backend development; make sure you remember my preference for using FastAPI over Flask."
2. "Remind me next Friday at 2 PM to review the quarterly budget projections."
3. "Based on our previous conversation about my morning routine, adjust your morning check-in to be more concise."

Tips & Limitations

• Tip: Always execute prune.sh at the end of every session to ensure that your active working memory is safely archived into your episodic layers, preventing data loss.
• Tip: Use snapshot.sh before major configuration changes to compress your memory layers and save local storage space.
• Limitation: Memory growth is cumulative; while the architecture is efficient, extremely long-term usage may eventually require manual pruning of low-priority sensory data to optimize latency. Always ensure the context.sh output is injected correctly, as failure to do so will result in the agent operating without its memory layer overrides.