Context is rent
AI does not have a memory problem. It has a diet problem.
Everyone is waiting for a bigger model to fix a habit problem. The fix is not more room — it is being deliberate about what the machine has to carry before you ask it anything. Here is the whole discipline, in plain words, with nothing held back.
The misdiagnosis
It is not forgetting. It is overeating.
More capacity does not fix a habit problem.
Everybody says AI has a memory problem, so everybody sits and waits for a bigger model to fix it. That is the wrong diagnosis, and waiting is the wrong treatment.
A machine's working memory is a fixed space. Everything you put in it — documents, history, instructions, whatever the last tool handed back — competes for the same room. Fill it carelessly and three things get worse at once: what it costs, how long it takes, and how often it is wrong.
More capacity does not fix that. A bigger window filled the same sloppy way degrades exactly the same way. Later, and for more money.
This is not a theory we like the sound of. It is the finding MIT published after examining three hundred real deployments: the failures were not about how clever the models were, they were about systems that could not retain anything, could not adapt to how a particular business works, and could not improve.
The discipline
What we actually do about it.
These are our own habits, run daily on our own work and measured rather than assumed. There is nothing proprietary in this list. We are publishing it because a business that adopts all six and never buys a thing from us is still better off, and because we would rather be judged on whether it works.
Summaries first, detail on demand
Never open the whole library to answer one question. An index, then a summary, then the one document that matters. Go deeper only when the work forces it.
One record, no copies
Duplicated truth gets loaded twice, contradicts itself eventually, and can never be safely thrown out. Keep one authoritative version and point everything else at it.
Write it down
So the next session reads a page instead of re-reading a week. What a machine knows tomorrow is whatever somebody wrote down today.
Scope the worker
A hand sent to do a small job carries a small briefing. Handing every worker the whole company is the fastest way to burn money and confuse the work.
Structure beats prose
Designed records compress meaning. Loose paragraphs do not. The difference shows up on the bill every single day.
Measure it or you are guessing
We watch our own usage and chase down the expensive shapes. The expensive shape is almost always searching — a machine hunting for something nobody prepared.
Context is rent. Every page you load pays it, every session, for as long as you keep loading it.
Findings
The six ways it goes wrong.
In roughly the order we run into them, out in real businesses.
Dumping everything in
Pasting the whole folder into the window is not context. It is noise with a hosting fee, and the machine attends worse, not better.
Re-reading instead of recording
No memory, so every session starts from nothing and works out yesterday again. The bill doubles and the answers drift.
Five copies of the truth
Five loads, four of them stale, and no way to clean up — because you cannot safely delete what you cannot prove is the wrong copy.
One giant do-everything assistant
It carries do-everything context. Ten small scoped hands beat one omniscient one on cost, speed, and being right.
Treating the bill as weather
Usage never measured, so waste is never seen, so the conclusion becomes "AI is expensive" instead of "we are wasteful."
Waiting for a smarter model
The next model inherits your habits. A better engine on a badly run operation is a badly run operation that costs more.
Payoff
This one survives the price war.
There are three ways out of a metered bill and two of them are shopping. Run the model on hardware you own. Rent raw compute and run it yourself. Both are real ways out, but both are purchasing decisions. They change who you pay and how.
The third is to need less. That is not a purchasing decision. It is a statement about how well a business had its own work in order before anyone put a machine near it, and it is the only one of the three that keeps paying after the prices settle, because it was never really about the price.
A system that cannot safely forget cannot reliably remember. Getting that right is ordinary good management, and people managed it before electricity.