How to Work Around AI Amnesia

I'm hardly the first person to talk about this. But I think it's one of the most essential - and most misunderstood - challenges in working with LLMs. And if you haven't hit it yet, you will.

Limited working memory. Technically, it's called the "context window." I call it AI amnesia - because that's exactly what it feels like.

When you start an LLM chat session - ChatGPT, Gemini, Claude, whatever your tool of choice - it's like sitting down with a brilliant person you've never met. They don't know you. They don't know your projects. They have no memory of previous conversations. The tools are starting to layer in persistent memory in various ways, but the baseline assumption still holds: every conversation starts at zero.

You can overcome it. You can give your LLM cumulative knowledge and build sessions that compound on each other. The method is surprisingly simple. Not too different, actually, from how you'd handle this with a human.

Let's say you just hired a new employee and you wanted them to go through a stack of invoices - pulling key data into a spreadsheet. What would you tell them? You'd explain what the invoices are for. You'd show them the spreadsheet layout. You'd point out exactly which data fields matter. You might walk through an example or two and ask if they had questions.

That's it. That's the entire playbook for working with an LLM. All of that briefing information? Write it in a text file. Literally bullet it out. Then paste it into your chat at the start of your session - or tell your LLM to read the file directly. Just like that context would empower a human to do the work, your LLM will now be equipped to do exactly the same.

Always instantiate your LLM session with the knowledge, context, and data it needs for the topic or task you're working on. Think of it as briefing a brilliant new collaborator before they touch anything.

There's a catch, though. You can overdo it. A lot of context at the start will give you wisdom at the cost of working memory. Using my analogy: the quick briefing I just described is enough. But what if you spent an hour briefing your new employee instead? Or handed them thirty pages of documentation before they touched a single invoice? At some point you've overloaded them - and the actual work never gets done.

Same thing happens with an LLM. Load too much context and the model starts losing the thread mid-session. Results get worse. The very detail you were trying to preserve gets crowded out.

• Treat your context file as a living document. Start lean. Add to it over time as you discover what your LLM actually needs to know.
• A lot of context at the beginning equals more informed judgment - but less headroom to operate. Too little equals more room to think, but a less prepared collaborator.
• Tell it what it needs to do the work, same as you'd do with any smart new hire.
• There is a balance. And like most things in knowledge work, the calibration depends on the task. But the general rule holds: enough to orient without overwhelming.

I've watched smart people dump massive briefing documents into LLM sessions and wonder why results got worse instead of better. I've also watched people start completely context-free every time and wonder why the LLM keeps missing the point. The answer, both times, is the briefing. This is one of those things that's invisible until you've experienced it - and obvious the moment you have.