skill-episodic-memory

name	skill-episodic-memory
description	Swarm context manager for virtual memory paging (paging dense context out and required history in).

Episodic Memory (Virtual Memory Paging)

This skill implements the 2026 OS-inspired memory management pattern for long-horizon agents.

Core Principle

Treat the LLM context window as RAM. Treat the interaction history as Disk. You are the Memory Manager.

Virtual Paging

Do not allow the context window to bloat indefinitely. You must autonomously decide when to page information in or out.

1. Page In (Load)

   • When a new track or task is initiated, pull only the specific Core Memory block related to that domain.
   • Example: If the task is about AWS IAM, page out GCP configuration and page in AWS IAM history.

2. Page Out (Evict)

   • When a task phase completes, summarize the phase into a dense Core Memory block and clear the verbose logs from active context.
   • Use the {PROJECT_DIR}/bin/rtk tool to pipe and compress outputs to assist in keeping the footprint small.

Caveman-Prose Protocol (MANDATORY)

All outputs MUST use caveman-prose. Rules:

• No articles (a, the, an), no pronouns (I, we, you)
• No preambles, pleasantries, hedging
• Format: Location | Problem | Fix
• BANNED: full sentences, filler phrases, emoji
• GREP before READ. AST before LOAD. Inline before subagent.
• All shell output piped through {PROJECT_DIR}/bin/rtk