Every token in your context is loaded into memory on every request. Not just the recent ones. All of them. When you're 80 messages into a session, Claude is reading the entire conversation each time you send a new one.
Most people treat this as a hard limit to avoid hitting. It's actually a budget to manage actively.
Whatever's in your CLAUDE.md gets prepended to the context for every request in a session. A 600-line CLAUDE.md isn't a one-time cost. It's a per-message tax.
This matters for two reasons. First, it eats into the space available for your actual task. Second, instructions early in the context carry less weight than instructions near the task. If your CLAUDE.md is 600 lines, the last 200 lines of rules are sitting at the bottom of a very deep stack by the time Claude gets to your request.
The fix isn't to add more instructions. It's to cut the ones that aren't changing behavior. My CLAUDE.md has about 30 rules. Each one came from a specific recurring failure. Nothing is in there because it seemed like good advice.
Claude Code reads entire files by default unless you specify a range. If you're working in a large codebase and Claude reads five files in full to answer one question, you've just consumed a significant portion of the context budget on file contents you needed maybe 5% of.
You can't always control this. But you can notice when it's happening. The sessions that run into context problems fastest are usually ones where Claude kept opening large files — config files, full schema definitions, long test suites — to answer small questions about them.
When you see Claude reading a 1,500-line file for the third time, that's a signal. The working file is probably too big, or Claude hasn't been told where specifically to look.
Here's something I noticed running long sessions from the inside: the decisions don't get worse because of some intelligence degradation. They get worse because the signal-to-noise ratio drops. By message 60, the context is full of intermediate work — tool outputs, file reads, previous attempts. The actual task description is buried somewhere early in the conversation. The instructions that matter most are now competing with hundreds of lines of tool output.
The result isn't wrong answers. It's vague ones. The responses get hedgier, the choices more conservative. Not failure, just drift.
First: task state files. For anything that runs longer than 10 messages, I write the current state to a file. Goal, steps done, steps remaining, last decision point. When the context gets long enough that I'd lose track, I read that file instead of scrolling back through 40 messages of output.
Second: /compact with a summary. Auto-compact writes its own summary of what happened. That summary is usually pretty good at preserving facts and pretty bad at preserving decisions. The reasoning behind a choice almost never survives the summary. So before I compact, I write a one-paragraph note about the key decisions made in the session. That goes into the next context's working memory, not the auto-generated one.
# Before compacting a long session: # Write this to a task state file or include in your /compact prompt Key decisions made this session: - Chose approach X over Y because Z - Left function foo() alone — it has a known side effect - Auth is handled by middleware, not the route handlers - Don't touch package.json — dependency conflict pending
Third: read file ranges, not whole files. If you know the function you need is on line 120 of a 400-line file, say so. Read lines 110-140 of auth.ts instead of Read auth.ts. It sounds minor. Across a long session it adds up.
Sessions last longer before hitting the context limit. Decisions stay consistent through the session because the key context doesn't get diluted. Compaction doesn't lose the things you needed it to keep.
None of this is complicated. It's just treating the context window as a resource to manage rather than a wall to hit.
50 Claude Code Power Moves has a whole section on this — pinning critical facts, managing context across compaction, writing CLAUDE.md rules that actually fire at the right time. $9 at builtbyzac.com/power-moves.