The reason AI "forgets" you has nothing to do with AI

Ever notice AI seems sharp early in a chat, then starts losing the thread the longer you talk? Or feels like a total stranger the next time you open it? That's not a glitch. It's something psychology has understood for decades, and once you see it, AI stops feeling unpredictable.

What context-dependent memory actually is

In a well known (if since debated) study, researchers had divers learn a list of words either underwater or on land, then tested their recall in both environments. The divers who learned underwater remembered significantly more when tested underwater than when tested on land, even though it was the exact same information. The environment itself became part of the memory. Recall wasn't just about what you knew. It was about whether the surrounding context matched the context you were in when you learned it.

The broader concept, environmental context-dependent memory, is well established, though the effect tends to be modest and not every attempt to replicate the original study has succeeded. The underlying principle has held up better than that one specific experiment.

AI context is context-dependent memory, running in real time

AI doesn't have a memory that persists independent of context, the way you might assume a computer would. It has something closer to what those divers had. Everything said earlier in a conversation is the environment. The model's ability to respond well depends entirely on whether that environment, that context, is still present when you need it.

Open a new conversation and you've changed environments completely, on land instead of underwater. Nothing was deleted. It's just not the environment where that information lives anymore. That's not a flaw in the system. It's the same principle behind why you can't always recall something outside the room you learned it in.

That's also why AI can feel like it drifts in a long conversation. The earlier parts don't disappear, they just get further from where the conversation is now, the same way details from the start of a long meeting get harder to recall than something said a minute ago.

Why this actually matters

• Keep important details close to where you need them used. Distance genuinely hurts recall, for AI and for people.
• Don't expect a new conversation to remember anything. Different environment, clean slate.
• If something from earlier still matters, just say it again. You're not being annoying. You're putting the cue back in the room.

This is also a good reminder that understanding context isn't optional anymore. It's part of using AI well. The people getting the most out of these tools aren't the ones with the fanciest prompts. They're the ones who understand what the tool can and can't hold onto, and work with that instead of against it.

One more thing worth knowing

This isn't fixed forever. AI is evolving fast, and how much context a model can hold, and how well it uses it, changes with almost every new release. The specifics here will keep shifting. The underlying principle, that recall depends on what's actually present, probably won't.

The tool is new. The principle behind it isn't.

Happy Learning,

KP