Cog gives Claude Code persistent memory, self-reflection, and foresight. No server, no runtime — just CLAUDE.md, markdown files, and slash commands.
It started as a personal experiment: what happens if you give Claude a structured memory system and let it audit its own rules? Over time it grew into a three-tier memory hierarchy (hot/warm/glacier), a self-improvement pipeline where Claude detects contradictions and condenses patterns, and a foresight layer that cross-references what it knows against what's coming.
Everything is plain text by design. You can observe exactly how the model organizes knowledge and where it fails. The architecture draws from RLM (recursive memory), Zettelkasten (thread framework), OpenViking (tiered context loading), and progressive summarisation.
When you connect MCP tools (calendar, email, GitHub), the memory layer gives them context they don't have alone — it stops being a note system and starts behaving like a cognitive layer.
marciopuga•1h ago
It started as a personal experiment: what happens if you give Claude a structured memory system and let it audit its own rules? Over time it grew into a three-tier memory hierarchy (hot/warm/glacier), a self-improvement pipeline where Claude detects contradictions and condenses patterns, and a foresight layer that cross-references what it knows against what's coming.
Everything is plain text by design. You can observe exactly how the model organizes knowledge and where it fails. The architecture draws from RLM (recursive memory), Zettelkasten (thread framework), OpenViking (tiered context loading), and progressive summarisation.
When you connect MCP tools (calendar, email, GitHub), the memory layer gives them context they don't have alone — it stops being a note system and starts behaving like a cognitive layer.
Clone it, run /setup, start talking.
Docs: https://lab.puga.com.br/cog