The Paradox
AI labs kept promising “infinite context” — one million tokens or more — to handle any level of complexity. The pitch sounded like an end to the problem of AI forgetting things. It was not. The more context you add, the less reliable the model becomes.
The trap is in a wrong assumption. We treat context as memory — a hard drive that stores state reliably. It is not. Context is a probabilistic attention surface. As it grows, the weight of your original instructions dilutes. The model does not forget your rules. It stops caring about them because they are drowning in recent noise.
The Failure Pattern: The 100k Cliff
Across sustained builds, I observed a consistent degradation arc. It is not random — it is structural.
From zero to 10,000 tokens, coherence is high and the model follows system prompts closely. From 10,000 to 100,000 tokens, execution stays stable but needs occasional correction nudges. Past 100,000 tokens, three things happen. Architectural rules set in the first exchange get ignored. Code that looks locally correct contradicts the project schema at the system level. And you start spending more tokens arguing with the model to follow the rules than you spend actually building.
At that point, the session is net-negative. Continuing does not just slow you down — it corrupts the work.
The Correction: The Stateless Session Model
To maintain reliable output, stop trusting session history. Treat every AI interaction as disposable and stateless.
The fix has three parts.
First, move state out of the AI and into a living document — a Google Doc you control. AI cannot tell the difference between an active constraint and an inactive one inside a long thread. A document you curate does not have that problem. It contains only the truths that are still current.
Second, do not let a session die on its own terms. Kill it deliberately. Before ending any session — or the moment you notice drift — run what I call the Migration Handoff. Paste this prompt into the session:
“Generate a Migration Document for the next instance. One: summarise current state. Two: list active constraints. Three: list the immediate next steps. Constraint: optimise this text so that when I paste it into a fresh AI session, it has full context to resume work immediately.”
Third, paste the Migration Block into your Google Doc and start a fresh session. Entropy resets to zero. The model is back to its best.
The Tradeoff: Reliability Over Convenience
This workflow is annoying. It destroys the illusion of a smooth, uninterrupted conversation. I think of it as the Junior Employee Tax — you do not just chat with a junior employee. You run standups. You review handovers. You manage context.
I accept this overhead because unreliable output is more expensive than context management. Every hour spent debugging contradictions in a long, rotting thread costs more than the five minutes it takes to do a Migration Handoff.
The Signal
Context is not memory. Context is friction. Until models can natively distinguish between active and inactive constraints, you must do that work for them. Short, stateless sprints beat long, drifting threads every time.