The smartest systems we've ever built share one catastrophic flaw: they don't know what to forget. And the more capable they get, the more expensive that blind spot becomes.
We've been watching this play out in three places at once, and the thread running through all of them should matter to you if you're building anything that's supposed to last.
Start with the machine
The latest AI research coming out of IBM and Hugging Face is asking a pointed question about AI agents: how much memory does an agent actually need? Not "how much can it hold" but "how much should it retain." The distinction sounds academic. It isn't. An agent that holds everything is not smarter than one that holds the right things. It's just slower, noisier, and more likely to act on stale context. The engineering problem of building useful AI agents turns out to be, at its core, a pruning problem. You have to design for forgetting, not just for remembering.
This is counterintuitive when everyone is racing to give models longer context windows and bigger retrieval systems. Nvidia, OpenAI, Anthropic, Google are all pouring billions into the infrastructure of accumulation. More data. More compute. More storage. The bet is that more memory equals more intelligence. But that's the wrong framing. A hospital with every record from 1970 is not smarter than one with a clean, current chart. It's just harder to navigate.
Now look at what happens when institutions forget how to prune
The US healthcare research apparatus is one of the most complex knowledge systems humans have ever assembled. Decades of longitudinal studies, clinical trials, population data, institutional memory about what works and what kills people. And right now, that system is being torn apart, not because it was pruned intelligently, but because no one built in a mechanism to do so. The research agency had accumulated so much bureaucratic weight, so many siloed programs, so many misaligned incentives, that it became politically vulnerable to someone who could show up and call all of it broken. When you never prune, someone else will eventually do it for you. And they won't be careful.
That's the institutional version of the same problem. The system grew without discipline. It held onto everything. It couldn't tell the difference between mission-critical knowledge and procedural bloat. So it looked like a target, and it became one.
And then there's the personal version
Alan Dershowitz, according to The Atlantic, is still relitigating a social snub from a few years ago. He has new friends, a new context, a new chapter. But the old grievance still runs in the background, eating cycles, shaping how he presents himself, coloring every conversation. He cannot let the old context window close. And so it costs him, in reputation, in energy, in the story he tells about himself to everyone he meets.
This is not a piece about Dershowitz. It's a piece about what happens when a person, a system, or a product cannot distinguish between what still matters and what's just noise that got saved.
The business version is the one you should lose sleep over
We see this constantly. A founder builds something that works. They accumulate processes, tools, vendors, habits, tribal knowledge, team rituals, pricing assumptions, positioning language, client relationships that should have ended two years ago. None of it gets pruned because none of it is obviously broken. It just quietly compounds into weight. By the time they feel it, the business is dragging. Revenue is there. Momentum isn't.
The irony is that the period of hardest accumulation usually follows the period of fastest growth. You sprint, you grab everything that helps you run faster, and then you stop sprinting and realize you're still carrying all of it. The tools that saved you at $500K are strangling you at $2M. The client you kept because they were your first is now consuming 30% of your service capacity at 15% of your margin. The internal process that someone built when you had three people doesn't scale to fifteen and nobody has the political capital to kill it.
This is a memory problem. Your business remembered everything and forgot to ask what was still worth keeping.
Pruning is a skill, not a one-time event
Good trail runners know this intuitively. The best loop routes are designed to bring you back to where you started with exactly what you need for the terrain ahead, nothing more. You don't carry your summit gear on the flat stretches. You don't carry your flatland pace on the climb. The loop is forgiving because every lap teaches you what to shed.
The best-run businesses operate the same way. They build in review cycles that are actually reviews, not just check-ins. They kill tools. They exit clients. They deprecate workflows. They do it on purpose, on a schedule, before someone else forces it on them. They treat context management as part of operations, not as a sign of instability.
What we've built into our own practice, and what we help clients build into theirs, is a simple habit: every quarter, you look at everything your business is holding and you ask two questions. Is this still earning its spot? And are we keeping it because it's useful, or because getting rid of it is uncomfortable?
Those two questions will surface more leverage than any new tool you could add.
The agent problem is your problem
When you hire an AI agent, or a contractor, or a new service layer, you are adding memory to your system. That memory will compound. The question isn't whether to add it. The question is whether you've built in the discipline to prune it when it stops earning its place.
The labs building the smartest AI systems in the world are now engineering for strategic forgetting. The healthcare institutions that didn't build that mechanism are being hollowed out. The smart operators we've worked with over the past 25 years all share one habit: they throw things out before they have to.
The thing slowing your business down is probably not something you need to add. It's something you need to cut, and you've been waiting for a better time to do it.
There isn't a better time. There's just now, and later when it's worse.