Why it happens
This is not a broken connection, and it is not specific to Hjarni. MCP gives an assistant a set of tools it may call. Whether it calls them on any given turn is the model's decision, made from the conversation so far. Ask something that sounds like general knowledge and the model will often answer directly, because nothing in the question signals that your private notes are relevant.
So the failure is quiet. Nothing errors. You just get a confident, generic answer, and the note you wrote last week sits unread.
Three layers, weakest to strongest
Hjarni already does the first two for you. The third is the one that reliably changes behaviour, and it takes about thirty seconds.
01 · automatic
The server asks on connect. When your client connects, Hjarni's MCP server sends session instructions telling the assistant to check your brain and read your saved instructions before working. How prominently a client surfaces that varies, and some ignore it, which is why it is not enough on its own.
02 · you write them once
Instructions inside Hjarni. You can write AI instructions for your whole brain, for your personal space, or for a single folder. A Runbooks folder can say "search this folder first and follow the steps exactly." These shape how your AI uses the notes it finds, and the assistant reads them when it checks in. They are powerful, but they still depend on it checking in.
03 · the one that works
One line in your client's own instructions. This sits in the assistant's system prompt, so it applies every turn rather than depending on the model choosing to look. Paste it into Claude project instructions, Claude custom instructions, a CLAUDE.md for Claude Code, or ChatGPT custom instructions:
Before answering questions about me, my work, or my projects,
search Hjarni first and answer from what you find there.
After we decide something worth keeping, save it as a note.How to tell whether it is searching
Every MCP client shows tool calls somewhere: an expandable "searched Hjarni" row in Claude, a connector indicator in ChatGPT. If you see no tool call, the answer came from the model alone, however plausible it sounds. The fastest test is to ask something only your notes could answer, such as a decision you recorded but never discussed in that chat. A generic answer means it did not look.
Why we do not claim this is automatic
Plenty of memory products imply the assistant simply knows everything you have saved. In practice every MCP-based tool shares this behaviour, because the protocol is the same and the decision belongs to the model. We would rather tell you where the seam is, and how to close it in one line, than have you discover it a week in and conclude the notes were not being read.
The trade you get for that seam is worth it: the notes are plain Markdown you can read and edit, and the same set is available to Claude, ChatGPT, Cursor, and anything else that speaks MCP. Nothing is extracted, summarised, or rewritten on your behalf.