The context window is your AI’s “working memory.” It’s the amount of text it can hold in its digital brain at one time. In 2026, we have models with massive windows—Gemini 1.5 Pro goes up to 1 million tokens—but there’s a catch: the “Middle Degradation” problem. At about 60-70% utilization, the model starts getting “stupid.” It forgets variables you mentioned ten minutes ago or re-introduces code you already told it to delete. If you’re still fighting your bot because it’s “lost,” you aren’t managing your context correctly.
Why does my AI start forgetting things halfway through?
It’s not a bug; it’s an attention budget. Models compress and deprioritize earlier parts of the conversation to make room for new input. You’ll see this when the bot starts contradicting itself or ignoring specific instructions from the start of the chat.
The Scenario: You’ve been debugging a messy auth issue for two hours. The chat is a mile long. You ask the model to add a simple log statement, and it writes something that uses a library you aren’t even using. The session is “drunk” on context. It’s time to purge.
Trick 1: Start a new session for every task
This is the single most effective thing you can do. Each new conversation starts fresh with full attention capacity. If you’re doing three things—debugging a function, writing a new feature, and updating docs—do them in three separate sessions.
The Scenario: You’re in the middle of refactoring a React component and you suddenly remember you need to fix a CSS bug in the footer. Don’t ask the current session to do it. Open a new tab, paste the footer CSS, fix it, and come back. Your main session stays “clean.”
Trick 2: Paste only what’s relevant
Stop pasting entire files when you only need help with one function. Cut out the noise. A 200-line file where the issue is in a 20-line function? Paste the function and the relevant types.
The Scenario: You’re frustrated because a specific API call is failing. You paste the whole
api.tsfile, thepackage.json, and theReadme.md. The model gets overwhelmed by the noise and misses the simple typo in line 42. Cut the crap and just paste the failing call.
Trick 3: Use a “Session Summary” to restart
When a session gets too long but you can’t lose the state, ask the model to summarize itself: “Summarize our progress, the key architectural decisions we made, and what the next step is. I need to move this to a fresh session.”
The Scenario: You’ve built a complex data pipeline over 50 messages. The bot is starting to lag. You get the summary, paste it into a new chat, and suddenly the bot is fast and accurate again. You’ve successfully “defragmented” its memory.
Trick 4: The “Selective Amnesia” prompt
You can tell the model to ignore certain parts of its context. It’s not a hard delete, but it refocuses the attention mechanism.
The Scenario: You spent 30 minutes talking about a database migration that you decided NOT to do. Tell the bot: “Forget everything we discussed about the Postgres migration. Focus only on the current Prisma schema.” This prevents the “ghost” of the old plan from haunting your current code.
Trick 5: Prefer Diffs over full file rewrites
When asking for edits, tell the model: “Only provide the specific lines that changed or a unified diff format.”
The Scenario: You have a 500-line file. You ask the bot to add one
ifstatement. If it rewrites the whole file, you’ve just used 2,000 tokens for a 10-token change. Over five edits, you’ve filled your context window with redundant code.
Trick 6: Use local context files like CLAUDE.md
If you use tools like Claude Code, keep a CONTEXT.md file in your repo. It should list the current state of the project and your coding preferences.
The Scenario: You’re tired of telling the AI every single time that you prefer using Tailwind for styling and tabs for indentation. Put it in
CONTEXT.md. The AI reads it at the start of every session without you having to re-paste it. It’s automated “onboarding” for your agent.
Trick 7: Know when to switch models
Not all tasks need a 1M token window.
- Gemini 1.5/2.0: Use for “Read Only” tasks like analyzing a massive codebase or 500-page PDF.
- Claude 3.5 Sonnet: Use for interactive coding and reasoning. It’s smarter within its 200k window than Gemini is at 800k.
The Scenario: You’re trying to find a security vulnerability across a massive monolithic repo. Use Gemini to “read” the whole thing once and find the problem. Then, take that specific snippet to Claude to actually write the fix.
FAQ
Why does the AI start hallucinating near the end of a session? Because its attention is spread too thin. It’s trying to satisfy too many past constraints at once and starts taking shortcuts or guessing.
Is it really better to have multiple short chats? Yes. Every time. Short chats are faster, cheaper, and 10x more accurate.
Want to understand how the brain of the AI actually works? Check out What Is a Context Window?.
