bulletproof-memory_skill

Overview

This skill implements a Write-Ahead Log (WAL) protocol plus a SESSION-STATE.md file to give agents reliable, persistent working memory across compaction, restarts, and mid-conversation distractions. It forces state writes on user input so context is saved before the agent responds. The result is an agent that recovers automatically and never loses active task state.

How this skill works

On every user-provided concrete detail, the agent updates SESSION-STATE.md before composing a reply. SESSION-STATE.md acts as hot RAM: current task, immediate context, key files, and last-updated timestamp. Startup and recovery protocols prioritize reading SESSION-STATE.md, then recent daily notes and memory search, so the agent can resume without asking the user to repeat itself.

When to use it

• When agents frequently lose context after compaction or restarts
• During long-running tasks that span multiple sessions
• When users provide corrections, decisions, or any concrete details
• If your agent asks "what were we doing?" or shows gaps in memory
• When context usage approaches system limits and you need reliable flushing

Best practices

• Trigger writes on user input only — update SESSION-STATE.md before responding
• Make SESSION-STATE.md the first thing read at session start
• Keep SESSION-STATE.md concise: current task, immediate context, key files, timestamp
• Implement a context-usage monitor and flush rules (e.g., >60% write before each reply)
• Use a recovery protocol that reads SESSION-STATE.md, daily notes, then memory search

Example use cases

• A planning assistant that must preserve decisions and action items across restarts
• Customer support agents that capture user corrections immediately to avoid repeated errors
• Developer bots working on multi-file tasks that need key file paths and current goals preserved
• Long-term coaching sessions where progress, corrections, and next steps must survive compaction
• Automated handoffs where a new session must pick up exactly where the last left off