Working Memory Is a Budget

I can always tell when we’ve overloaded a new hire. It’s not something they say. It’s something you see. It’s a look that’s half determination, half drowning. They’re nodding in meetings, taking notes furiously, saying “that makes sense” when it clearly doesn’t yet. And the hard part is, they’re doing it because they want to prove they belong. The eagerness and the overwhelm are fighting each other, and overwhelm usually wins quietly.

This is why I like to create “chunks” when onboarding people to one of my teams. Two weeks. A month. Ninety days. Not because people can’t handle the information eventually, but because they can’t handle it all at once. I give every new hire a “Working with Kelsey” guide before they start — not a company manual, just a few pages about what it’s like to work with me, what I expect, how I communicate. It helps them understand the immediate environment before they try to map the entire landscape.

The principle behind this isn’t complicated: working memory is finite. Every person, every team, every system has a limited amount of information it can actively hold and use at one time. When you exceed that limit, you don’t get someone who knows more. You get someone who can’t find anything.

The Same Lesson at Every Scale

I wrote about this concept in a newsletter issue in the past. It covered the seven leadership lessons I’d learned from training AI. One of those lessons was about overfitting. Overfitting is when we flood a model with too much data and instead of learning useful patterns, it starts memorizing noise. In data science we have ways to plan for and minimize this. Unfortunately, in human organizations we don’t always see it, let alone manage it. I experienced the same thing happening with a team that was drowning in dashboards, tracking seventeen metrics when three would have told them everything they needed to know. More information does not always lead to better results.

That pattern keeps showing up. I see it when I onboard new team members. I see it when we introduce a sufficiently complex project to a team — and it shows up every time I watch someone set up their AI tools.

Here’s what I mean. When most people start working seriously with AI, they do the same thing new hires do. They want to prove they can get value from it, so they front-load everything. Upload fifty documents. Paste in every brand guideline, case study, client brief, and process document they have. The instinct makes sense: give it more information, get better results.

But AI has working memory too. It’s called a context window, and it works almost exactly like human working memory. Everything the AI can “see” and think about in a single conversation has a hard limit. When you fill most of that space with background documents, there’s less room for the actual thinking. That means your actual questions, the back-and-forth, the work you’re trying to produce get less of the AI’s brain.

When I loaded eighty past client case studies into Claude, every conversation felt sluggish. Responses were generic. Claude used to hit limits, but now it compacts your conversation. That’s good because you can continue, but every time that conversation gets compacted you lose some of the signal. The fix was simple: I moved the case studies somewhere Claude could search them when needed and kept a five-page summary of the key patterns in the active workspace. The same information was available, but the result was completely different. Because the working memory wasn’t drowning anymore.

Three Questions That Work for People and Machines

The three simple questions can fix this for AI, for new hires, and for teams are always the same:

1. Does this person or system need this information right now, or just sometimes? Some of the things they reference constantly belong in active memory. While other things they reference occasionally belong somewhere accessible but not loaded by default. My “Working with Kelsey” guide is active memory for a new hire. The full employee handbook is reference material they’ll search when they need it.
2. How much is it? There’s a point where volume defeats purpose. Twenty pages of core context that someone actually uses beats two hundred pages that sit unread. This is true for onboarding documents, team wikis, and AI workspaces equally.
3. Is it stable or constantly changing? Stable information like your values, your methodology, your communication preferences can be loaded once and trusted. Information that changes frequently like project status, market data, or customer feedback — should be pulled fresh when needed, not baked into the foundation.

With Claude specifically, this maps to three tools. Projects can hold the stable core: who you are, how you work, what you always need the AI to know. MCPs let Claude search larger archives on demand. Things like your past work, research databases, anything too large or too fluid for active memory. And Skills give Claude repeatable processes or workflows that combine multiple sources and tools in a consistent way.

The Instinct That Keeps Getting Us

But the tools are just the current implementation. They’ll change. Claude Code has already shifted how I think about some of this. The way you structure context for an agent driven system is different from how you structure it for a conversational assistant, but the underlying principle is identical. You’re still managing a finite budget of attention. You’re still deciding what belongs in working memory versus what belongs in reference.

The principle doesn’t change: curate ruthlessly, organize around how the user actually thinks, and build for the question they’re asking right now rather than every question they might ever ask.

This is what I’ve learned watching it play out at every scale. The new hire who gets a focused two-week onboarding ramps faster than the one who gets a 200-page wiki and a “good luck.” The team that has three clear metrics outperforms the one tracking seventeen dashboards. The AI workspace with ten pages of curated context produces better work than the one with a hundred and fifty pages of everything-just-in-case.

The instinct is always to add more. More documentation. More context. More information. It feels like diligence. It feels like preparation. But every system with finite working memory — human or artificial — hits the same wall. Past a certain point, more information doesn’t produce better understanding. It produces noise.

The new hire who’s drowning in information isn’t failing because they’re not smart enough. They’re failing because someone spent the budget before they walked in the door. Same thing happens with AI. Same thing happens with teams.

The fix is always the same. Don’t give them everything. Give them what they need to do the work in front of them, and make the rest findable.