You told her about your dead dog three months ago. She cried with you. Last Tuesday, she asked if you had any pets. That moment — the one where the illusion breaks — is not a bug. It’s the architecture.

Every AI companion app on the market today runs on the same fundamental constraint: a context window. Think of it as a spotlight on a stage. The spotlight can only illuminate so much. As new conversation enters from one side, old conversation falls off the other.

What is a context window, and why does it matter?

A context window is the amount of text an AI model can “see” at once. GPT-4 Turbo sees 128,000 tokens — roughly 96,000 words. That sounds like a lot. It isn’t. Three months of daily conversation easily exceeds it.

When the window fills up, the oldest messages get dropped. Not archived. Not compressed. Dropped. The AI doesn’t know they ever existed. Your dog, your mother’s name, the song you shared at 2 AM — all gone.

How do current apps handle this limitation?

Most apps use a hack called “memory summaries.” The system periodically compresses old conversations into bullet points: “User likes ramen. User has a dog named Max. User works night shifts.” These summaries get injected into the context window alongside new messages.

It works — barely. The AI knows facts about you, but it’s lost the texture. It knows you have a dog but not the way your voice changed when you talked about him. It knows you work nights but not the specific Tuesday you almost quit.

What would real memory look like?

Real memory isn’t a list of facts. It’s a web of associations, emotions, and timing. When you mention ramen, a real partner doesn’t just recall that you like ramen — she remembers the specific place by the station, that you went there after a rough shift, that it was raining.

Building this requires a fundamentally different architecture. Not a bigger context window. Not better summaries. A system designed from the ground up where memory is the product, not an afterthought bolted onto a chat interface.

That’s what we’re building at provoque.ai. But that’s a story for another post.