Why Your AI Sounds Different at Hour 3 Than It Did at Hour 1

You start a project with your AI in the morning. The drafts are sharp. The voice matches yours. It remembers the constraints you gave it. Three hours later, something is off. The tone has slipped. It’s ignoring rules you set up early. It feels like a different assistant.

You’re not imagining it. The AI is different now. Here’s what’s happening, and what to do about it.

The short answer: things scrolled out of view

An AI model can only “see” a fixed amount of text at once. That limit is called the context window. Everything in the conversation — your messages, the AI’s replies, any documents you’ve pasted, the instructions you gave at the start — counts toward that limit.

When the conversation runs long enough, the oldest stuff falls out. The model can no longer reference it. The careful tone guide you wrote at the top? Gone. The example outputs you said to copy? Gone. Whatever’s still visible is what’s shaping the next reply.

From the AI’s side, it isn’t forgetting on purpose. It literally can’t see what it can’t see.

Why it feels like personality drift

Once your tone rules scroll out, the AI defaults back to its baseline voice. That baseline is fine, but it’s not your voice. The shift feels personal because the output sounds different. It isn’t personal. It’s just the model running without your style sheet anymore.

The same thing happens with constraints. You said “never link to Twitter” an hour ago. Three hours later, links to Twitter. You said “keep paragraphs short.” Now they’re long. None of that is the AI ignoring you. It’s the AI no longer being told.

Three concrete moves that fix it

1. Put the rules in a file the AI re-reads

The simplest fix is to stop relying on the AI’s short-term memory at all. Write your tone rules, project context, and constraints into a file the AI loads at the start of every reply. In Claude Code, that file is called CLAUDE.md. Other agentic tools have equivalents (instructions files, system prompts, project briefs). The principle is the same: rules live on disk, not in conversation.

The deeper version of this pattern is in The CLAUDE.md Pattern — one file per project, loaded automatically, never scrolls out.

2. Start a fresh session for new work

If you’ve been chatting for two hours and now want to draft something important, open a fresh chat. Counter-intuitive but true: a new session with your full briefing pasted in is sharper than an old session that’s already eaten through its context window.

The same principle applies if you’re using a system that automatically “compacts” old conversation when context fills up. Compaction summarizes what came before, but summary is lossy. The first thing that gets crushed is nuance — which is exactly the part you tuned for.

3. Split work across separate channels or projects

If you’re jumping between three different projects in one conversation, all three sets of context have to share the window. None of them get the AI’s full attention. Run each project as its own conversation, or its own channel, or its own folder — whichever your tool supports.

This is one reason a vault-and-channel setup beats one giant chat. The whole approach is laid out in Obsidian as a Second Brain: The Claude Code Setup.

When hour 3 matters and when it doesn’t

If you’re asking your AI a single question, none of this matters. One question, one answer, done. Drift only shows up in long-running, high-stakes work: drafting a chapter, researching a launch, building a campaign over hours.

For one-shot tasks, ignore everything in this post. For long projects, set up the rules file, get comfortable opening fresh sessions, and treat the context window the way you’d treat working memory in your own head — finite, valuable, easy to clutter.

The mental model that makes the rest easy

An AI session is a whiteboard, not a notebook. Whiteboards run out of room. Whatever’s on the board now is what shapes the next decision; whatever was erased is gone. If something matters across the whole project, it doesn’t belong on the whiteboard. It belongs in a file the whiteboard re-reads from.

Once that switch flips in your head, your AI stops drifting. Not because the model got better. Because you stopped asking it to remember things it can’t.

---

Read next: The CLAUDE.md Pattern for the structural fix · Your AI Workflow Is One Spilled Wine Away From Gone for what else can quietly fail.