Context engineering is not just for AI teams. It is what your notes should do for you.

There is a phrase AI engineers have started saying a lot: context engineering. It sounds like a job title. It is really a simple idea.

A model can only see so much at once. Its context window is finite. So the work is not cramming everything in. It is putting the right information in front of the model at the right moment, and leaving the rest out until it is needed. Anthropic's own teams describe the best version of this as just-in-time: the agent does not load everything up front, it pulls the specific context it needs, on demand.

Here is the part nobody says out loud. When you sit down and chat with Claude or ChatGPT, you are doing context engineering by hand. Badly. Over and over.

You are the context window

Think about how a normal session starts.

You open a fresh chat. The AI knows nothing about you. So you paste in your stack. You explain the project. You remind it how you like things formatted, what the constraints are, what you decided last week and why. Then you ask your actual question.

Next session, the window is empty again. You do it all over. You are hand-feeding the model its context, one paste at a time, starting from zero every single time.

That is the human bottleneck. The models got huge context windows, and agent tooling got just-in-time retrieval. You got copy and paste. Your AI forgets everything between sessions, so the burden of remembering falls on you, and you carry it manually.

The fix is not to paste faster. It is to stop being the retrieval system.

Curation is context engineering

Here is the move. Write the recurring context down once, as notes, in a place your AI can read. Then let the AI pull what it needs, just in time, instead of you pushing it every time.

That is the same just-in-time idea the engineers are chasing, pointed at your own work. The context lives outside the conversation. The assistant retrieves the relevant pieces when the question calls for them. You stop re-explaining yourself because the explanation is already written down and reachable.

This is what a knowledge base is for. Not a place to hoard notes you never open. A place that feeds your AI the right context on demand.

Three things make it work.

Write down the context that repeats

If you have explained it twice, write it once. Your architecture decisions. Your customer interviews. Your style rules. The reasons behind the choices you will be asked about again. Anything you keep re-typing into a chat is a note waiting to be written.

The test is simple. The next time you start to paste a paragraph of background, stop and save it instead.

Make the notes the AI can actually use

A note that helps you is not automatically a note that helps an AI. The model needs structure: a clear title, a real summary, the context stated plainly instead of implied. We wrote a whole piece on this, because it is the difference between notes that get retrieved and notes that get skipped. See your notes are not agent-legible yet.

Good curation is half of context engineering. The other half is retrieval.

Let MCP do the retrieval

This is the just-in-time part. A knowledge base with a built-in MCP server lets Claude or ChatGPT search your notes and read only the ones relevant to the moment. The assistant is not carrying your entire knowledge base in its context. It is pulling the two notes that matter for this question, when the question is asked.

That is the whole loop. You author the context. MCP retrieves it on demand. The conversation stays focused, and you stay out of the copy-paste business. For how to structure the notes so retrieval lands well, best practices for an MCP wiki is the practical guide.

The shift this asks for

It is a small change in habit with a large payoff. You stop treating each chat as a blank slate you have to brief. You start treating your notes as the memory your AI reads from.

Generation stopped being the hard part a while ago. The models write fine. The hard part is getting the right context to them at the right time, which is exactly why retrieval is where the work moved. We made that case in full in generation is solved, finding it again is the problem.

Context engineering for a model is the engineer's job. Context engineering for yourself is just writing things down where your AI can reach them.

Give your AI the context once

Stop re-explaining your project to your AI every session. Write the context once. Let the AI pull it just in time.

Hjarni is a knowledge base with a built-in MCP server. You write notes in Markdown. Claude and ChatGPT search and read them, follow your instructions, and remember what you told them across every conversation. Write once. You both remember.

For the concept behind it, AI long-term memory lays out how reusable context across conversations actually works. When you are ready to wire it up, connecting Claude or ChatGPT takes about five minutes.

Agent tooling already moved to just-in-time memory. This is how you get it too.