A Brain in Five Dimensions: How We Keep a Memory That Doesn't Rot

Ask most people how an AI "remembers" and you'll get one of two answers: it's in the prompt, or it's in a vector database. Both are real. Both are also a single shape, and a single shape can only answer a single kind of question well.

We kept running into the limits of that. The way you want to retrieve a fact depends entirely on why you're reaching for it — and "why" changes constantly. So we stopped picking one storage shape and started storing the same knowledge in several at once.

One fact, five questions

Take a single thing the system knows. Depending on the moment, you might want it answered in completely different ways:

• Tell me exactly what we decided — you want the precise record, verbatim, unblurred.
• What's related to this? — you want meaning-based neighbors, the stuff that's "about the same thing" even if it shares no words.
• How does this connect to that? — you want the relationships, the graph of what-touches-what.
• What happened, in order? — you want the timeline, the sequence that turns facts into a story.
• What do I need right now, given what I'm doing? — you want the slice that's relevant to the task in front of you, and nothing else.

Those are five different questions, and no single database answers all five well. A keyword store nails the first and fumbles the second. A vector store nails the second and can't do the third. A graph nails the third and ignores time. So we keep all of them — the same knowledge, written five ways, each tuned to one of those questions. When the system reaches for something, it reaches into the shape that fits the need.

Why this beats "just use a vector database"

The fashionable answer to AI memory is to embed everything and search by similarity. It's genuinely useful — it's one of our five. But lean on it alone and you get a system that's good at "what's vaguely like this?" and bad at everything else. It can't tell you the exact wording of a decision, can't trace a chain of consequences, can't reconstruct the order things happened.

Memory that only does similarity is memory that slowly turns everything into a warm blur. You want the blur and the sharp edges. You want to feel your way to the right neighborhood and then read the exact street sign. Five dimensions instead of one is how you get both.

Packing the whole situation to go

There's a sixth idea that rides on top, and it's the one I'm fondest of. Sometimes you don't want to retrieve a fact — you want to capture an entire situation and be able to set it down again later, intact. Where we were, what we'd decided, what was still open, what to do next.

We call those capsules, and they're how a thread of work survives being put down and picked back up — by us, or by a fresh instance of the AI with none of the original context. You pack the situation into a portable bundle, and "remembering" becomes "unpacking," not "reconstructing from scratch." It's the difference between a colleague who read your handoff note and one who's guessing what you meant.

The honest version

None of this is magic, and I won't pretend it's finished — keeping five views of the same knowledge in sync is real work, and we get it wrong sometimes. But the principle has earned its place: an AI's memory should match the shape of the question, and questions come in more than one shape. Store accordingly, and the memory stops rotting into mush. Store one way, and no matter how clever the model, it's eventually remembering everything as a vibe.