Cog — Cognitive Architecture for Claude Code

Cog is an experimental cognitive architecture layer for Claude Code that adds persistent memory across sessions. Instead of starting fresh each conversation, it uses plain text files and markdown-defined rules to store and organize knowledge. A nightly pipeline consolidates conversations, extracts patterns, prunes stale data, and rewrites rules — analogous to REM sleep. Memory is organized in a three-tier system (hot/warm/glacier) with domain-based indexing. The project is designed as a learning tool to observe how AI models organize knowledge and whether self-reflection improves behavior over time.