mmag_skill

Overview

This skill implements Mixed Memory-Augmented Generation (MMAG) to give agents layered, persistent memory. It organizes five cognitive layers—conversational, long-term user, episodic, sensory, and working—so agents can remember across sessions, personalize responses, and stay context-aware. Use it to inject a prioritized system prompt that reflects current and historical context.

How this skill works

At session start you run context.sh and inject its output as the system prompt; that block is a unified, prioritized view of all memory layers. During a session use store.sh to append entries to the appropriate layer (long-term, episodic, sensory, conversational, working). Background scripts prune, snapshot, encrypt, and retrieve memories, and context.sh decrypts transparently in-memory when needed.

When to use it

• When you need agents that retain user preferences between sessions
• When tracking scheduled events, deadlines, or follow-ups
• When adapting responses to location, time of day, or environment
• When maintaining coherent multi-turn conversations across sessions
• Before running a model call to ensure the system prompt reflects memory state

Best practices

• Always invoke context.sh at session start and inject its output as the system prompt
• Store concise, labeled facts for long-term memory and timestamped data for episodic entries
• Prune working memory at session end to archive relevant items into episodic memory
• Encrypt long-term personal data and back up the encryption key securely
• Use retrieve.sh to audit stored items regularly and delete unnecessary or incorrect entries

Example use cases

• A personal assistant that remembers user scheduling preferences and upcoming meetings
• A support bot that recalls user account preferences and previous troubleshooting steps
• A location-aware agent that adjusts recommendations based on current weather or time
• A coaching bot that tracks progress across sessions and recalls milestones
• An automated workflow agent that snapshots working memory before long operations