Why an AI second brain breaks when your team uses it
Obsidian plus Claude, Notion AI, ChatGPT memory: what people in forums love about AI second brains, what frustrates them, and what changes for a company.
The AI second brain is one of the most discussed setups of the year. Connect your notes app to Claude or ChatGPT, often over MCP, and suddenly years of notes answer back. Forums for Obsidian, Notion and the AI assistants are full of setups, screenshots and, increasingly, complaints.
We read through a lot of those threads, because we build a company brain and wanted to know where the personal version works and where it breaks. Short version: what goes wrong for one person goes wrong much faster for a team.
What people like about it
The most common praise is surprisingly modest: the notes finally get used. Collecting was never the hard part; finding the right note again was. With an assistant on top, the archive becomes something you can ask.
The second thing people value is not having to explain themselves every time. In a Hacker News discussion comparing how ChatGPT and Claude remember, one user put it plainly: “Memory is by far the best feature in ChatGPT and it is the only reason I keep using it.” They didn’t want to give every new chat the context of their project again.
Both points carry straight over to companies. People want the AI to already know how things work here, without pasting it in each time.
What frustrates people
The graveyard. The most repeated complaint in note-taking forums is some version of: capturing works, maintaining doesn’t. Setups get abandoned after a few weeks, or turn into an archive that only gets written to. Adding AI makes the pile searchable. It doesn’t make it tidy.
Memory you don’t control. The same user who praised ChatGPT’s memory added: “But I am not a big fan of the conversation memory created in background that I have no control over.” Another wrote that there is “some slop from conversations I might have no desire to ever refer to again.” An AI that remembers everything also remembers the half-thought, the test, the thing you changed your mind about.
Rules that get ignored. Developers keep instruction files like CLAUDE.md or AGENTS.md so their coding agents follow house rules. A typical complaint: “claude routinely ignores stuff from CLAUDE.md, even with warning bells etc.” A long file of everything you care about competes with itself.
One file per tool. When Claude Code started reading AGENTS.md in September, one reaction was relief: “Maybe I can finally delete that useless symlink to the actual AGENTS.md.” People had been keeping the same rules in several files, one per tool, and linking them together by hand.
Where it breaks for a team
Those are annoyances for one person. In a company, they turn into real problems.
Stale becomes wrong. In a Hacker News thread about a “company brain” startup, one comment named the issue precisely: “context that’s right on Monday is quietly wrong by Friday, and nobody notices until a decision gets made on it.” In a personal vault, an outdated note costs you a minute. In a company, the AI repeats it to a customer.
Contradictions need a decision, not a guess. In the same thread someone asked what happens when sources contradict each other. The founders’ honest answer: for now, “always trust humans, and trust recent human info more than old human info”, with more sophistication planned. Newer is a reasonable default for your own notes. In a company, the newest statement is often a sales rep’s assumption, and the older one is the policy. So the useful question is who decides, and the date doesn’t answer that.
Nobody’s notes are the company’s notes. A personal second brain has one owner, so the question of what counts never comes up. Put five people’s notes behind one assistant and the AI can’t tell the draft from the decision, or the intern’s guess from the CFO’s rule.
Memory per tool is a copy per tool. ChatGPT remembers one thing, Copilot another, the coding agent reads a third file. Each is a private copy that drifts on its own.
What to keep from the personal version
The second brain crowd got a few things right, and a company setup should keep them:
- Capture has to be easy. If adding something takes effort, people stop. That’s why we collect company rules by voice interview rather than documentation projects.
- Plain, small units. Notes that do one thing are easier to find and to fix. For a company, that unit is a rule with a reason.
- A live connection. The setups people are happiest with let the AI read the source directly, not a copy.
What a company has to add
Three things a personal second brain never needed:
- An Owner for each area, who decides what counts. Recency doesn’t.
- Conflicts that surface instead of being quietly resolved by whichever document the AI found first.
- One approved version for every tool, so ChatGPT, Copilot and the coding agent read the same rule, not three copies of it.
Less goes in, on purpose. A company brain holds what someone approved, not everything anyone ever wrote down. That’s the difference we built CtxCore Brain around: a second brain for the company, but one that decides what applies.
If you want to see how that plays out in practice, we wrote down what we learned running one on ourselves.