7M Seven-Layer Memory System

Problem

Long agent sessions often fail because context grows too fast. Large read, test, or build outputs can consume the window before useful state is recorded, forcing expensive summarization or losing user preferences and subagent results. 7M treats memory as a layered defense, letting low-cost actions block higher-cost ones.

How It Works

The skill activates seven layers by trigger:
- L1 Tool Result Storage: writes large outputs to disk and keeps a preview, reducing context pressure.
- L2 Micro Compact: cleans stale tool results when idle, without calling an LLM.
- L3 Session Memory: maintains incremental task state across multi-turn work.
- L4 Full Compact: summarizes context when overflow occurs and session memory is unavailable.
- L5 Auto Memory Extraction: captures user feedback, preferences, and patterns.
- L6 Dream: consolidates memory in the background during idle periods.
- L7 Cross-Agent Communication: shares state for subagents and multi-agent workflows.

Boundaries

This is a memory and context-management policy, not a complete agent framework. Full Compact depends on model summarization, and Cross-Agent adds coordination overhead. It is most useful for long programming tasks, multi-turn sessions, subagent collaboration, and setups where prompt-cache preservation matters.