On 20 July my Claude Code 5-hour limit was gone after a single chat, on the Max plan. I assumed a bug. It wasn't. Here is the actual audit I ran, the four checks that found it, and the real before-and-after numbers from my own machine.
I want to save you the two days I lost chasing these.
Instruction files and memory files are not read once. They are attached to the context and re-sent on every message you type. A big one multiplies across an entire chat.
# from your project root
wc -c CLAUDE.md MEMORY.md AGENTS.md .cursorrules 2>/dev/null
# mine, before:
544533 MEMORY.md
1275 CLAUDE.md/usage in Claude Code. It shows the 5-hour and the
weekly bar separately. They are different limits and they fail for different reasons.This single change did almost all of my saving. The memory file becomes a table of contents; the write-ups move into separate docs that only get read when that area is actually being worked on.
# MEMORY.md keeps one line per topic:
- [Android app](docs/order-android-app.md) — build + install gotchas
- [Delivery flow](docs/delivery-flow.md) — order → dispatch → POD
# the 10 KB write-up itself lives in docs/ and is read on demanddocs/archive/. You will almost never need them again, and they cost you on every message
until you move them.I was maintaining a 65 KB list of my installed skills and tools inside my memory file. The tool already injects that list itself, every session, for free. I was buying it twice.
Each connected plugin or MCP server contributes tool names, schemas and instructions to your context. Disabling one unused plugin removed 30 skills from my baseline.
| What | Before | After |
|---|---|---|
| Memory file | 532 KB (3,671 lines) | 24 KB (341 lines) |
| Tokens from that file, per chat | ~140,000 | ~6,000 |
| Injected skills | 113 | 83 |
| Total starting context | ~157,000 | ~23,000 |
These are my measured numbers on my machine, not a promise about yours. Your saving depends entirely on how bloated your own files are. If your memory file is already 5 KB, checks 1 and 2 will do nothing for you and check 4 is where to look.
No. Cursor, Copilot, Windsurf and the rest all have some
always-loaded instruction file. The names differ — CLAUDE.md, AGENTS.md,
.cursorrules — the arithmetic is identical.
It cut my starting context by about 85%. I have no idea what it will do for you, and anyone who promises you a number without seeing your files is guessing. Run check 1 first — that measurement tells you whether the rest is even worth your time.
Not at all, and I'm not claiming that. People have genuinely hit limits faster at several points in 2026 and the vendor has acknowledged it publicly before. My point is narrower: in my case I blamed the tool twice and both times it was my own file. Check yourself first, because that's the part you can actually fix today.
It helps, and it's why a bloated file is survivable for a while. It is not free, it doesn't cover every turn, and it stops hiding the problem once the file gets big enough. Mine got big enough.