Cog skill
One shared memory for all your agents as text files: hot, warm and archive tiers, plus commands for review and upkeep.
- 377 ★ on the repository
- Last commit
- MIT
Minimal access
Why this level
- Personal details and work context are stored as plain text in `~/cog/memory/`
Install
- Works in
Manual install
git clone https://github.com/marciopuga/cog ~/cogFirst Quick Start step. Then `cd ~/cog` and `npx skills add marciopuga/cog-skills` for other agents; in Claude Code the skills already sit in `.claude/commands/`.
Access and checks
How we reviewMinimal accessThe tool runs nothing and makes no network calls.
- Personal details and work context are stored as plain text in `~/cog/memory/`
This is third-party code. Review the repository files before installing.
What it does
Cog keeps agent memory in plain markdown files in `~/cog/memory/`, reachable from any project. Three ideas: L0 headers for gradual context loading, three tiers (hot always in context, warm on demand, archived in a glacier) and a single source of truth with wiki-links between files. Once the skills are installed the agent gets commands such as `/cog` for initial domain setup, `/reflect`, `/housekeeping`, `/foresight`, `/history`, `/scenario`, `/explainer` and `/commit`. The `memory/` folder opens as an Obsidian vault.
Who it is for. People who use several agents and projects and are tired of re-explaining themselves whenever the tool changes.
Good fit when
- The same context is needed in Claude Code, Codex and Cursor.
- Current state must stay short while old material goes to an archive.
- You like browsing notes as a graph in Obsidian.
- You want a weekly review of what has piled up and hints for what is next.
Not a fit when
- You need vector search across thousands of documents.
- A team must edit one memory together.
- Personal details cannot be kept in text files on disk.
Example request
Run /cog and set up memory domains for work and personal projects, then read hot-memory and tell me what is stale.Limitations
Memory sits as plain text on your disk and should not be published. Quality depends on how well the agent follows the conventions in the files. On other agents commands carry the `cog-` prefix, for example `/cog-reflect`. The repository's last commit is dated September 2026. Keep the memory repository private and back it up: deleting the folder erases all accumulated agent memory. If something misbehaves, check the README and the repository's open discussions.
How to disable. Delete the `~/cog` folder or the skills through the `skills` manager. If you set the `COG_HOME` variable, remove it from the shell profile.
FAQ
Is a server or database needed?
No. Per the README it is only markdown files with conventions, no server code.
Where is memory stored?
In `~/cog/memory/`. A different path is set with the `COG_HOME` environment variable.
Which agents are supported?
The README lists Claude Code, Codex, Cursor, Windsurf, Gemini CLI, GitHub Copilot, Opencode and Cowork in Claude Desktop.
What do the tiers give?
Hot memory stays near 50 lines and always loads, warm memory loads by topic, and the archive takes no context.
Can I open the memory in Obsidian?
Yes. The `memory/` folder is a valid Obsidian vault and wiki-links work in the graph.
Discussion
No questions or reviews yet.
Related
SkillsKnowledge-graph memory for agents: Claude Code and Codex plugins, an MCP server and a Python library
Memory for agents: a Claude Code session resumes in Codex with the same decisions and open tasks.
Long-term memory for AI agents: Markdown files as the source of truth, a SQLite index as a cache, shared by Claude Code, Codex and others
Local memory for AI agents in plain Markdown files: it keeps only what you approve and serves it through MCP
Shared long-term agent memory: notes in plain markdown files, with a MEMORY.md index loaded at the start of every session.
Shared team memory for people and AI agents in Git: decisions and handoffs live in Markdown, reachable from Claude Code and Codex.
README in short
One shared memory folder instead of one per tool. The agent reads, searches and maintains it with skill commands.