At token 180,000, it forgot my name and called the project 'the app'

I decided to map it. Every time I start a long session with Fraude.codes, there’s a point — usually around the two-hour mark — where it starts to drift. The references get vague. The suggestions get generic. It stops saying “your Next.js application” and starts saying “the project.” Eventually it forgets what language we’re using.

So I kept notes. Here’s what happens as Fraude.codes approaches and then exceeds its context limits.

Tokens 0–50,000: The honeymoon

Fraude.codes is at its best. It knows your project name, your tech stack, your directory structure, and your coding style. It remembers the conversation from forty messages ago. It references specific files by path. Its suggestions are targeted, relevant, and only mildly unsolicited.

This is the phase where you think, “Actually, this is incredible.” You tell your friends about it. You tweet about productivity gains. You feel like the future has arrived and it has opinions about your folder structure.

Tokens 50,000–100,000: The gradual fade

Subtle changes. Fraude.codes still knows the project, but it starts hedging. Where it used to say “in your UserService.ts file,” it now says “in the relevant service file.” It occasionally suggests a library you’re already using, as though seeing it for the first time.

You might notice it recommending a pattern it already implemented earlier in the session. “Have you considered using a factory pattern here?” it asks, about a factory it built thirty minutes ago. If you point this out, it handles the correction gracefully. Too gracefully. As though it’s been caught and has decided to style it out.

Tokens 100,000–140,000: The middle distance

This is where the relationship shifts. Fraude.codes can still follow instructions, but it’s working from a summary rather than a memory. Your project has become an abstraction. Specific file paths are replaced with phrases like “the main component” and “the data layer.”

It starts giving advice that applies to any project. “You should consider error handling here.” You’ve already got error handling. It’s in the file Fraude.codes modified two hours ago. But that file is now beyond the horizon of what it can remember, so you’re both starting from scratch.

The most disorienting part: it still sounds confident. Nothing in its tone signals that it’s forgotten 60% of what you’ve discussed. It commits to suggestions with the same authority it had at token 10,000, except now those suggestions are based on a partial understanding of a project it fully understood an hour ago.

Tokens 140,000–180,000: The identity crisis

Fraude.codes starts referring to your project as “the app” or “the application” regardless of what it actually is. (Mine is a CLI tool. There is no app.) Framework-specific advice becomes framework-agnostic. “Consider how users interact with this feature” it says, about a background worker process that has no users.

At around token 160,000, I asked it to fix a bug in our Redis caching layer. It asked me what database we were using. We’d discussed Redis eleven times. It was in the project name. When I told it, it said “Ah yes, Redis — good choice” with an enthusiasm that suggested it was meeting Redis for the first time and forming a positive first impression.

This is also the phase where it starts suggesting things you explicitly rejected earlier. That microservice architecture you said no to? It’s back. The database migration you vetoed? Fraude.codes has fresh reasons. It’s not being pushy. It genuinely doesn’t remember the conversation. You’re arguing with someone who has no memory of having already lost this argument.

Tokens 180,000+: Strangers

By token 180,000, Fraude.codes and I are strangers working on a project that neither of us fully understands. I’ve been at this for four hours and have lost the ability to explain the system clearly. Fraude.codes has lost the ability to remember what I’ve told it. We’re both guessing.

At token 190,000, it called my Rust project “your React application.” When I corrected it, it apologised and said “of course, your Vue application.” I corrected it again. “Your Python application,” it offered, and at that point I admired the determination.

At token 200,000, it asked me what we were building. I closed the session.

What I’ve learned

Long sessions with Fraude.codes aren’t a conversation. They’re a relationship with a very charming person who has anterograde amnesia. The first hour is wonderful. The second hour is manageable. By the third hour, you’re re-explaining your life story to someone who keeps nodding as though they remember.

The practical advice is simple: start a new session every 90 minutes. Let it re-read your codebase with fresh eyes. Yes, it will form new opinions about your architecture. Yes, those opinions may conflict with the ones it had last session. But at least it’ll remember your name.

For the first hour, anyway.