From the free Starter Kit · file 5 of 5, shown in full

The seven ways this system rots

Read this after your first win, while you still like the system.

The hard thing about an AI knowledge base isn't building it. You just did that in twenty minutes. The hard thing is that an unmaintained knowledge base doesn't stay neutral. It goes wrong, in predictable ways, and it usually starts around week three. I've built and run these systems for years, and every dead one I've seen died from one of these seven.

I'm naming them so you recognize them early. What to do about each one is a maintenance discipline, not a paragraph. That's the honest boundary of a free kit.

01The junk drawer

Adding feels productive, so everything goes in and nothing comes out. Six weeks later the base is big and the AI's answers are mushier than in week one. Old facts and current facts carry equal weight, the AI can't tell which is which, and neither can you. Why it happens: capture has a dopamine hit; pruning doesn't. Every system drifts append-only unless something actively pushes back.

02Stale facts in a confident voice

You raised your prices in October. The BUSINESS doc still says the old number, so your AI quotes it, fluently and confidently, maybe in a draft you almost sent. Why it happens: documents don't know they're outdated. Reality changed, the doc didn't, and nothing in the system noticed the gap. The other rot modes waste your time. This one sends a wrong price to a client.

03The week-three abandonment

Setup energy is real but it isn't a routine. You built it on a motivated Saturday; nothing in your week tells you to touch it again, so the base freezes at its founding moment. Two months later it describes a business you no longer run, and using it feels like reading an old CV. Why it happens: the system was created by enthusiasm, and enthusiasm is not a schedule. You've probably lived this before with a note app, just without the AI attached.

04Instructions the AI quietly ignores

Every time the AI does something you dislike, you add a rule. The instructions grow, start contradicting each other, and the AI begins following them approximately. That reads as "the AI got worse" when actually the rulebook did. Why it happens: instructions get patched at the moment of irritation and never re-read as a whole.

05The shadow memory

You keep telling the AI important things in chats. A new client, a changed scope, a decision. None of it flows back into the base. The chat knew; the system never learned. The base diverges from reality one conversation at a time. Why it happens: in the moment, the chat working is all you need. The update step is nobody's job. The proposal rules in your agent instructions exist to catch exactly this, and they only catch it while you're still reviewing.

06The one-tool drift

Bit by bit, everything ends up inside one vendor's project, one vendor's memory, one vendor's format. The day you want to switch tools, or the tool changes under you, you discover your "portable source of truth" isn't. Why it happens: the path of least resistance always leads into the tool you use most. Portability is a property you maintain, not one you set up.

07The trust collapse

One day the AI gives you a wrong answer and you trace it back to the base. A stale fact, maybe, or a shadow-memory gap. From then on you double-check everything it says, and a memory you have to double-check is slower than no memory at all. So you quietly stop using it. Why it happens: the other six modes all end here. Trust in this system is binary, and it breaks on the first traced failure.

---

The pattern

None of these is a setup mistake. Your setup is fine. All seven are maintenance failures, and they're why "I built a knowledge base for my AI" and "my AI runs on a knowledge base I trust" are different claims separated by a discipline: update routines, review workflows, consistency checks, and a bounded weekly rhythm that keeps the whole thing alive without becoming a second job.

That discipline is the fourth block of Compound Context. Blocks one through three exist so there's something worth keeping alive. The founding edition comes with homework feedback on your real system and a 30-day unconditional guarantee: see the course.

And if you'd rather build the discipline yourself, that's genuinely fine. You now know the seven failure modes, which is more than most people who start this. Watch for stale facts and the shadow memory first; they arrive earliest.