10 Repos Can Piece Together Claude's Second Brain, but No One Will Clean Up the Contradictions for You

Ryven posted a list on X of 10 open-source repos that can already piece together Claude's second brain. His question is direct: why are we still building from scratch?

The background here is the LLM Wiki pattern proposed by Andrej Karpathy in April 2026. The process in the public summary is not complicated: put raw materials into a directory, have the model write them into a readable wiki, then do ingest, query, and lint based on the wiki. Ryven's long post makes the implementation steps more concrete: Obsidian handles viewing, Claude Code handles reading and writing, folders are split into raw/, wiki/, and output/, and rules go in CLAUDE.md. For ordinary users, this can be understood as: materials stay local, the model organizes them into interlinked pages, and when answering questions it tries to look only in those pages.

Ryven also included a caveat. At the end he notes that this walkthrough has not been run end to end, and menu names may change with app versions. This caveat is more useful than the list itself. Second-brain tutorials can easily look as if they'll work right after installation, but the real dividing line is the first ingest, the first permission prompt, and the first contradiction cleanup.

The 10 repos do not serve the same purpose.

• obsidian-skills: Agent Skills maintained by Obsidian CEO kepano, teaching Claude to write Obsidian Markdown, Bases, and Canvas, and also covering Obsidian CLI and web article body extraction. Suitable for people who have already decided to use Obsidian and want to first solidify the formatting and toolchain.
• claude-obsidian: Turns sources in an inbox into linked notes with citations. The repo page emphasizes transactional writes, hash validation, stopping on conflicts, and rollback. It addresses what to do when the model writes something wrong.
• obsidian-second-brain: Cross-session memory and a command collection. The tweet says 48 commands, the repo page says 49; the calendar command is available only in Claude Code, and builds for other platforms have 48. It also provides contradiction reconciliation, cross-source synthesis, and scheduled agents.
• karpathy-llm-wiki: Implements ingest, citation-backed querying, and lint around the Karpathy pattern, supporting Claude Code, Cursor, Codex, and OpenCode.
• second-brain: Closer to the minimal path: raw sources go in, linked wiki pages come out, and you browse them in Obsidian.
• second-brain-os: A suite-style project with a 77-page guide, a starter vault, 18 skills, 72 commands, and 6 sub-agents, 4 of which are read-only. Suitable for people who want a complete skeleton in one go.
• llm-wiki: Two-layer memory: each session loads 10 to 20 rules, and 50 to 200 wiki pages are pulled on demand. It cares about context budget.
• vouch: Changes are proposed first and then approved; claims without sources are treated as errors outright. Suitable for workflows with citation and approval requirements.
• ai-second-brain: Turns ChatGPT and Claude history into a linked Obsidian vault, which then grows into a wiki. It solves the problem of old conversations being scattered.
• LLM_Wiki_SecondBrain: The entire pipeline runs on local models without cloud APIs, with 32GB RAM recommended. Suitable for people who care about data not leaving their machine.

The list solves whether the wheels exist. It does not solve which one you'll still open a week later. Several replies on X are worth reading as a reality check. One user said that building a second brain may itself be a hobby, and organizing sources is just a cover story. Another said most projects die because the wiki grows faster than it is cleaned up, and asked which of these 10 actually handles contradictions. Someone else pointed the problem at writing: all 10 repos are solving reading, but what really rots is writing, because the agent does not know which detail today will still matter three sessions later; if no one writes it down, the index will grow around a hole.

These comments are closer to daily reality than the number of repos. Ryven's long post also includes corresponding maintenance actions: every week check for pages that contradict each other, pages without links, concepts that recur but have no standalone page, and claims without sources. It also mentions an easily overlooked permissions issue: writing 'Do not delete files' in CLAUDE.md is only a request; Anthropic docs say CLAUDE.md is context, not enforced configuration. To actually block deletion, you need permissions or a PreToolUse hook. This distinction is worth remembering. Rules written into a prompt and rules enforced by the system are two different things.

If the goal is simply to use it, don't start with 10 repos. Pick one, drop in 5 to 10 sources, ask a cross-source question, and save the good answer back into the wiki. First see whether it can give cited answers on your own materials, then look at maintenance cost. The first step in Ryven's original post is also at this scale; it is not about dumping in hundreds of files right away. Usage limits are shared across Pro and Max, and the first large ingest will consume them quickly.

If the goal is to enjoy building, then a different logic applies. Some people are willing to spend a weekend tuning the pipeline and figure out along the way which step breaks most easily. Just don't mistake building for using. A second brain truly starts working on the day you open it, ask a cross-source question, and the answer contains something you saved months ago but had forgotten. At that point, the number of repos does not matter; whether you can keep cleaning it up does.