I told ChatGPT I like short answers. Claude never got the memo.

Blog · 4 min read

Every AI tool you use is quietly building its own private idea of who you are, and none of them talk to each other. Tell ChatGPT you prefer short answers over long ones, and it might actually remember — next time, in ChatGPT. Open Claude for the same task five minutes later, and you're explaining it again, because as far as Claude's memory is concerned, that conversation with ChatGPT never happened. You didn't forget to tell Claude. You told a company, not a person.

That's the real shape of the problem, and it's not any one product's fault. OpenAI's memory stays in OpenAI's product. Anthropic's memory, when it has one, stays in Anthropic's. Every assistant's "it remembers you now" feature is scoped to its own walls, because nobody outside that one company has any reason to build the layer that lets it travel. The feature works. It's just stuck.

The workaround I'd settled for

For a while my fix was a text file: a running list of things I was tired of repeating, pasted into whichever tool I happened to be using that day — the ChatGPT tab if I was there, a Claude project's custom instructions if I'd switched. It worked exactly as well as you'd expect from something that only helps the tool you remember to paste it into. Switch to a third tool and the list wasn't there at all. Update one copy and forget the other, and now they disagree about which preferences are even current.

What changed

Skilnu memories live on your account, not inside any one assistant's memory system. Any MCP client you've connected — Claude, ChatGPT, Codex, whichever you're actually using that day — can save one with save_memory. Any MCP client you connect later, including ones you haven't tried yet, can find it again with search_memories. Tell ChatGPT once that you like short answers, and the next time you're in Claude, it's already there to find — not because Claude remembered, but because neither of them had to.

Retrieval is semantic, so the two don't even need to use the same words. A memory you phrased to ChatGPT as "please don't nitpick variable names" still turns up when Claude asks how you like your code reviewed, because the two sentences mean the same thing even though they share almost no words in common.

What this isn't

It doesn't replay your conversations, and it doesn't mean one assistant can read what you said to another — what crosses between ChatGPT and Claude is only what got explicitly saved, never the surrounding conversation. Nothing is stored unless something, you or the agent, calls save_memory, the same way nothing gets published as a skill until someone decides it's worth keeping. This is closer to a note you'd leave a new teammate than a transcript — a short, deliberate "worth knowing" list, not a recording.

And it needs the same discipline a skill does. "User prefers concise answers" is a memory worth having, in any tool. "Had a long conversation about databases on Tuesday" isn't — it's a summary with nothing to act on, the same failure mode a skill drafted from a document falls into when it describes instead of instructs. The useful memories read like instructions for next time, not like a diary entry about this time.