Maintainer huytieu landed COG-second-brain on GitHub's daily trending list on 16 November, with 1,486 stars and 37 new ones arriving that day S¹. The Python project pitches itself as a "self-evolving second brain" built on a V-model verification lifecycle where AI workers are barred from validating their own output, but every feature claim is self-reported with no independent testing cited S¹. The star count is a GitHub platform metric; no error bars, benchmark suite, or released test code accompany it, and the 1,486 figure measures repository popularity rather than deployed usage S¹.

My read: The V-model idea, borrowed from aerospace and embedded software, is the one genuinely interesting design choice here. Separating the agent that does the work from the agent that checks it is exactly the kind of discipline the agent ecosystem needs. What I don't buy yet is the scale. The title says 35 skills; the system diagram says 43. Neither number has been verified outside the maintainer's own README — these are the vendor's own figures with no third-party audit, and 1,486 stars measures curiosity, not production use.

The project's core claim is structural rather than algorithmic. COG stands for Cognition plus Obsidian plus Git, and it stores everything as markdown files under version control, with no database and no vendor lock-in S¹. The AGENTS.md file defines commands that AI agents can read and execute, and says it is compatible with OpenAI agents, Claude, and any AI that reads markdown P⁴. The README lists compatibility with Claude Code, Cursor, Kiro, Gemini CLI, and OpenAI Codex S¹.

Separating builders from checkers is COG's structural bet

The V-model is a systems-engineering approach where every development phase has a matching test phase. COG applies this to AI agents: six workers produce output, and four read-only verifiers check it, according to the README's system diagram S¹. The maintainer describes this as a closed-loop harness and instructs users to run ./scripts/validate-agent-surface.sh before publishing or updating framework files S¹. No released code or benchmark data demonstrates that the four verifiers actually catch errors at any stated rate; the claim rests on the maintainer's README alone S¹. The latest release, v3.16.0 on 23 October, added seven skills for post-shipping content including release notes, documentation, and announcement posts P².

The source document itself flags that the README is the maintainers' own description, not an independent assessment S¹. The skill count is inconsistent: 35 in the title versus 43 in the diagram. Terms like "self-evolving" lack technical definitions in the available material.

For a developer evaluating this today, the markdown-only architecture is the easiest thing to verify. Clone the repository, read the AGENTS.md file, and check whether ./scripts/validate-agent-surface.sh runs cleanly on your machine. A developer already using Claude Code or Cursor could test a single skill without abandoning their current setup S¹. A knowledge-work team lead running Claude Code today could drop COG's AGENTS.md into an existing repository and run the validation script to see whether the four read-only verifiers surface errors their current single-agent workflow misses S¹.

GitHub's release history shows active development through late October, with v3.16.0 shipping on 23 October P². The maintainer has not yet reconciled the 35-versus-43 skill count or published task-completion data for the verifier pipeline.


Sources: S1 — huytieu/COG-second-brain: Self-evolving second brain with 35 AI skills · P2 — Releases · huytieu/COG-second-brain · GitHub · P3 — Trending repositories on GitHub today · GitHub · P4 — AGENTS.md · P5 — icip-cas/COG


Written from 5 sourced items, 4 of them primary.

More from Not A Tech Guy