AI Doesn't Lie. It Forgets.

Every model has a limit to how much it can hold in mind at once — a context window. Think of it as short-term memory. Once a conversation runs long enough, the earliest parts fall off the edge to make room for what's new.

Here's the part that actually causes problems: when that happens, the model doesn't say "I don't remember that." It fills the gap with its best guess. Often the guess is wrong. That's what gets labeled a hallucination, and it's easy to read as the model being unreliable in general when really it just lost the thread.

What this looks like in practice

Say you're a few hours into a long coding session with an AI assistant. At some point you see a note like "context getting full — compacting conversation." That means the earlier turns are being summarized to free up space. If you set an important constraint at the very start of the session — "always use async/await," "this project uses Node 20," whatever it was — that detail might be gone or flattened into something vaguer by the time it matters again.

The assistant keeps working. It just isn't working from the instruction you think it still has.

That's not a bug in the tool. It's the tradeoff of a finite context window, and it applies to every model, not one specific product.

What actually helps

• Start a new conversation for a new topic rather than letting one thread carry everything from the last three unrelated tasks.
• Re-state constraints that matter, especially once a session has run long — a one-line reminder costs nothing and closes the gap.
• Break large tasks into smaller, focused ones. A clean, short context produces more reliable output than a long one carrying accumulated noise.
• Know when compaction is happening. Most coding assistants surface it directly — that notice is useful information, not just a status update to ignore.

None of this makes the model smarter. It makes you a better user of the memory it actually has, which in practice matters more than which model you're using.