marciopuga/cog

What it solves

Cog addresses the problem of siloed AI agent memory. Instead of each tool or project having its own isolated memory, Cog provides a centralized, plain-text memory layer that is shared across all AI agents and projects, preventing the need to re-explain context to every new tool.

How it works

Cog uses a structured system of markdown files and conventions rather than a database or server. It employs a tiered memory architecture:

• Tiers: Memory is divided into hot (always loaded), warm (on demand), and glacier (archived) tiers.
• L0 Headers: Agents use progressive context loading, scanning summaries (L0 headers) before reading full files.
• Single Source of Truth: Facts are stored in one place and cross-referenced using wiki-links.
• Agent Integration: It installs "skills" (instruction files) into various AI agents (like Claude Code, Cursor, Windsurf, and GitHub Copilot) to teach them how to read, search, and maintain this shared memory folder.

Who it’s for

Users who work with multiple AI agents across different projects and want a persistent, shared knowledge base that the agents can maintain and the user can manually edit via tools like Obsidian.

Highlights

• Agent Agnostic: Works across a wide range of supported agents including Claude Code, Codex, Cursor, and Gemini CLI.
• Human-Readable: Memory is stored as plain markdown files, making it compatible with Obsidian for visualization and manual editing.
• Observable: No black-box database; the entire memory system is observable via simple file system reads and grep.
• Automated Maintenance: Includes skills for weekly housekeeping, reflection, and strategic foresight to prune and consolidate memory.
• Recursive Loading: Uses a ladder of small reads to efficiently navigate large memory trees without overloading the agent's context window.